LOT 1 - Optimizare script build -Instalare mono comanda

2026-06-27 06:42:02 -07:00 · 2026-06-27 06:42:02 -07:00 · 42ff22bf85
commit 42ff22bf85
parent 5380c3fc63
127 changed files with 16163 additions and 532 deletions
--- a/ai_platform/local_gpu_stack/.env.example
+++ b/ai_platform/local_gpu_stack/.env.example
@ -0,0 +1,43 @@
+# =============================================================================
+# DiDi LOT 1 — Local GPU stack configuration
+# Copy to `.env` (deploy.sh face asta automat daca lipseste) si ajusteaza.
+# =============================================================================
+
+# Host IP folosit in EXTERNAL_URL-uri si in comenzile de test.
+# Lasa GOL -> deploy.sh detecteaza automat IP-ul principal si il scrie aici.
+HOST_IP=
+
+# --- Brain / Atomic / Postgres -----------------------------------------------
+# Parola e folosita de postgres, atomic-server SI brain-api (trebuie sa fie una singura).
+POSTGRES_PASSWORD=brain_local_dev
+# Token API atomic-server. Lasa GOL -> deploy.sh il genereaza automat
+# (atomic-server token create) si il scrie inapoi aici.
+ATOMIC_TOKEN=
+
+# --- Gateway -----------------------------------------------------------------
+GATEWAY_API_TOKEN=didi-local-dev-token-123
+
+# --- Dashboard ---------------------------------------------------------------
+DASHBOARD_DB_USER=dashboard
+DASHBOARD_DB_PASSWORD=devpass123
+DASHBOARD_DB_NAME=dashboard
+# staging=true -> bypass Keycloak (intri direct in UI). Scoate in productie.
+DASHBOARD_STAGING_MODE=true
+
+# --- Modele (nume servite; nu schimba fara motiv) ----------------------------
+LLM_MODEL=Qwen/Qwen3.5-35B-A3B
+LLM_SERVED_NAME=qwen3.5
+EMB_MODEL=BAAI/bge-m3
+RERANK_MODEL=BAAI/bge-reranker-v2-m3
+WHISPER_MODEL=large-v3-turbo
+WHISPER_REPO=mobiuslabsgmbh/faster-whisper-large-v3-turbo
+BUSTER_MODEL=l8cv/BusterX_plusplus
+# Extractoare (NER) — GLiNER + backbone-ul lui
+NER_MODEL=urchade/gliner_multi-v2.1
+NER_BACKBONE=microsoft/mdeberta-v3-base
+
+# Zona LOCALA unica pentru TOATE modelele (un folder per model, format HF cache:
+# models--<org>--<name>/). Toate containerele (Qwen/BusterX/bge/whisper/GLiNER)
+# sunt servite OFFLINE de aici. Lasa GOL -> deploy.sh foloseste ai_platform/models.
+# Ce lipseste se descarca automat din HF in acest director (offline-first).
+MODELS_DIR=
--- a/ai_platform/local_gpu_stack/README.md
+++ b/ai_platform/local_gpu_stack/README.md
@ -0,0 +1,74 @@
+# DiDi LOT 1 — Local GPU stack (deploy progresiv)
+
+Deploy „cărămidă cu cărămidă" al întregii platforme pe un host cu GPU.
+Pentru fiecare componentă: **build/pull + download → seed (dacă e cazul) →
+TEST funcțional → doar dacă testul trece corect → următoarea**. Idempotent,
+se oprește la primul test picat.
+
+## Cerințe (verificate de preflight, nu se instalează automat)
+- Docker Engine 24+ cu plugin `docker compose` v2
+- NVIDIA driver ≥535 + `nvidia-container-toolkit` (runtime configurat în Docker)
+- ≥1 GPU (proiectat pe 2× H200; GPU 0 = LLM+Whisper, GPU 1 = BusterX+bge×2)
+- ~100 GB liberi pe `/` (toate modelele ≈ 94 GB)
+
+## Modele — o singură zonă locală
+Toate modelele stau în **`ai_platform/models/`** (format HF cache, un folder per model),
+servite **offline** de acolo (fără rețea la runtime). `MODELS_DIR` în `.env` (auto = acest folder).
+Ce lipsește se descarcă o singură dată în el (`deploy.sh` pasul DOWNLOAD). NU se comite în git.
+
+| Folder | Model | Folosit de |
+|---|---|---|
+| `models--Qwen--Qwen3.5-35B-A3B` | LLM (67G) | llm (vLLM) |
+| `models--l8cv--BusterX_plusplus` | deepfake (16G, Qwen2.5-VL intern) | video (vLLM) |
+| `models--BAAI--bge-m3` | embeddings | embeddings (vLLM) |
+| `models--BAAI--bge-reranker-v2-m3` | reranker | rerank (vLLM) |
+| `models--mobiuslabsgmbh--faster-whisper-large-v3-turbo` | STT | audio |
+| `models--urchade--gliner_multi-v2.1` + `models--microsoft--mdeberta-v3-base` | NER + backbone | extractors |
+
+Copiezi `ai_platform/models/` pe altă mașină → zero download la deploy.
+
+## Utilizare
+```bash
+cd ai_platform/local_gpu_stack
+./deploy.sh                 # tot, de la preflight la dashboard
+./deploy.sh --from web      # reia de la cărămida 'web' (sare peste cele dinainte)
+./deploy.sh --only llm      # rulează o singură cărămidă (presupune deps deja up)
+./deploy.sh --install-deps  # instalează Docker + nvidia-container-toolkit (sudo), apoi continuă
+```
+`.env` se creează automat din `.env.example` la prima rulare.
+
+### Mașină nouă (VM „chioară" cu GPU)
+1. (o singură dată) driver NVIDIA ≥535 + reboot — `sudo ubuntu-drivers install` (nu se automatizează, cere reboot).
+2. `git clone <repo> && cd .../ai_platform/local_gpu_stack`
+3. `./deploy.sh --install-deps`  → instalează Docker+toolkit, apoi:
+   - **HOST_IP** se auto-detectează și se scrie în `.env`
+   - **BusterX** se descarcă automat din HF dacă nu e pe disc (`BUSTER_HF_CACHE`)
+   - Qwen/bge se descarcă (cu progres), totul se ridică cărămidă cu cărămidă cu test la fiecare.
+
+> Singurul pas care NU e o comandă: driverul NVIDIA (reboot). Restul e `./deploy.sh`.
+Cărămizi, în ordine: `embeddings · rerank · llm · extractors · audio · video · forensic · web · catalog · gateway · brain · dashboard`.
+
+La final: dashboard pe `http://<HOST_IP>:51300/admin-ai/` (staging mode, fără login).
+
+## Ce automatizează (fixurile care înainte erau manuale)
+- rețea unică `didi-network`; plasare GPU 0/1; nume containere `didiAI-*`
+- base-URL vLLM **fără** `/v1` (codul îl adaugă) pentru llm/embeddings/video
+- YOLO scrie modelul în `/tmp` (cwd e read-only la non-root)
+- gateway LLM injectează `enable_thinking:false` → JSON curat din Qwen3.5 (sentiment/OCR/semantic/brain)
+- SearXNG fără proxy outbound, ca serviciu compose
+- brain: aliniere parolă postgres, creare **token atomic** (auto, salvat în `.env`), **seed taxonomie** (79 tag-uri)
+- dashboard: `.env` cu staging + override-uri health; **seed catalog** (9 intrări)
+
+## Fișiere
+| Fișier | Rol |
+|---|---|
+| `deploy.sh` | orchestratorul progresiv gated |
+| `lib.sh` | logging, `wait_health`, `gate`, shim docker (merge și prin `sg docker`) |
+| `docker-compose.yml` | stack-ul GPU + non-GPU + searxng |
+| `.env.example` | configurația (host, porturi, parole, token) |
+| `seed/seed_catalog.sh` | populează catalogul DB al dashboard-ului (idempotent) |
+| `seed/seed_brain.sh` | token atomic + taxonomie + env brain |
+| `TESTING.md` | comenzi de testare per componentă + smoke-test |
+
+> Brain rulează din `modules/didi_brain/infra/docker-compose.yml`; dashboard și gateway
+> din `modules/*/deploy/` — `deploy.sh` le orchestrează pe toate sub proiectul `local_gpu_stack`.
--- a/ai_platform/local_gpu_stack/TESTING.md
+++ b/ai_platform/local_gpu_stack/TESTING.md
@ -0,0 +1,228 @@
+# DiDi LOT 1 — Comenzi de testare per componentă
+
+Host: **10.11.10.18**. Toate comenzile sunt `curl` copy-paste. Statusul din dreptul
+fiecăreia e rezultatul verificat la ultima rulare.
+
+> Convenție: `H=10.11.10.18`. Rulează întâi `export H=10.11.10.18`.
+
+---
+
+## 0. Pregătire fișiere de test (o singură dată)
+
+```bash
+export H=10.11.10.18
+cd /home/topclossers/didi-lot1-ai/ai_platform/local_gpu_stack
+mkdir -p test_assets
+# video mic din artefacte
+cp "$(find ../../artefacte_lot1/runs_deepfake -iname '*.mp4' -printf '%s\t%p\n' | sort -n | head -1 | cut -f2)" test_assets/sample.mp4
+# extrage un cadru (jpg) + audio (wav) cu ffmpeg din containerul extractors
+docker cp test_assets/sample.mp4 didiAI-extractors:/tmp/s.mp4
+docker exec didiAI-extractors sh -c 'ffmpeg -y -i /tmp/s.mp4 -frames:v 1 /tmp/frame.jpg -vn -ar 16000 -ac 1 -t 8 /tmp/audio.wav'
+docker cp didiAI-extractors:/tmp/frame.jpg test_assets/frame.jpg
+docker cp didiAI-extractors:/tmp/audio.wav test_assets/audio.wav
+cd test_assets   # restul comenzilor presupun că ești aici pentru fișiere
+```
+
+---
+
+## 1. LLM — Qwen3.5-35B-A3B  (port 14011)  ✅
+
+```bash
+# modele servite
+curl -s http://$H:14011/v1/models | python3 -m json.tool
+
+# chat completion
+curl -s http://$H:14011/v1/chat/completions -H 'Content-Type: application/json' -d '{
+  "model":"qwen3.5",
+  "messages":[{"role":"user","content":"Ce este un deepfake? Raspunde scurt."}],
+  "max_tokens":200
+}' | python3 -m json.tool
+
+# health
+curl -s http://$H:14011/health
+```
+> Notă: Qwen3.5 e model de **reasoning** — răspunsul conține un bloc „Thinking" înainte de
+> răspunsul final. Normal.
+
+## 2. Embeddings — bge-m3  (port 14100)  ✅
+
+```bash
+curl -s http://$H:14100/v1/embeddings -H 'Content-Type: application/json' -d '{
+  "model":"bge-m3","input":"text de analizat"
+}' | python3 -c "import sys,json;print('dim:',len(json.load(sys.stdin)['data'][0]['embedding']))"
+```
+> Așteptat: `dim: 1024`.
+
+## 3. Rerank — bge-reranker-v2-m3  (port 14200)  ✅
+
+```bash
+curl -s http://$H:14200/v1/rerank -H 'Content-Type: application/json' -d '{
+  "model":"bge-reranker-v2-m3",
+  "query":"detectie deepfake",
+  "documents":["o pisica neagra","sistem de detectie deepfake pentru video"]
+}' | python3 -m json.tool
+```
+> Așteptat: documentul relevant primește scor mult mai mare (index 1).
+
+## 4. Audio / Whisper — large-v3-turbo  (port 54300)  ✅
+
+```bash
+curl -s http://$H:54300/v1/audio/transcriptions \
+  -F "file=@audio.wav" -F "model=large-v3-turbo" | python3 -m json.tool
+# modele / info
+curl -s http://$H:54300/v1/models | python3 -m json.tool
+```
+> Acceptă și `.mp4` direct (ffmpeg în container). Returnează `text` + `language`.
+
+## 5. Video / BusterX++ (deepfake)  (port 54600)  ✅
+
+```bash
+# verdict deepfake REAL/FAKE/UNCERTAIN
+curl -s http://$H:54600/analyze/video -F "file=@sample.mp4" | python3 -m json.tool
+
+# analiza semantica (descriere continut, foloseste si LLM-ul)
+curl -s http://$H:54600/analyze/video/semantic -F "file=@sample.mp4" | python3 -m json.tool
+
+curl -s http://$H:54600/v1/info | python3 -m json.tool
+```
+> Așteptat: `{"verdict":"REAL"|"FAKE"|"UNCERTAIN","frames_analyzed":16,"evidence":[...]}`.
+
+## 6. Extractors  (port 54400)
+
+```bash
+# 6a. metadata: EXIF/ELA/ffprobe/integritate/C2PA  ✅
+curl -s http://$H:54400/v1/metadata -F "file=@sample.mp4" | python3 -m json.tool
+
+# 6b. NER (GLiNER, multilingv)  ✅
+curl -s http://$H:54400/v1/ner -H 'Content-Type: application/json' -d '{
+  "text":"Klaus Iohannis s-a intalnit cu Emmanuel Macron la Bucuresti."
+}' | python3 -m json.tool
+
+# 6c. detectie obiecte (YOLOv8)  ✅
+curl -s http://$H:54400/v1/detect -F "file=@frame.jpg" | python3 -m json.tool
+
+# 6d. OCR (LLM vision)  ✅
+curl -s http://$H:54400/v1/ocr -F "file=@frame.jpg" | python3 -m json.tool
+
+# 6e. sentiment (deleaga LLM)  ✅
+curl -s http://$H:54400/v1/sentiment -H 'Content-Type: application/json' -d '{
+  "text":"Produsul este excelent, sunt foarte multumit!"
+}' | python3 -m json.tool
+```
+> Așteptat: `{"ok":true,"results":{"label":"pozitiv","score":0.9,"rationale":"..."}}`.
+> (Gateway-ul LLM injectează `enable_thinking:false` pe vLLM → JSON curat din modelul de reasoning.)
+
+## 7. Forensic — rPPG / lip-sync / forgery / lighting  (port 8085)  ✅
+
+```bash
+# module disponibile
+curl -s http://$H:8085/api/forensic-modules | python3 -m json.tool
+
+# analiza forensica (multipart; campul fisierului se numeste 'video')
+curl -s http://$H:8085/api/forensic-evidence \
+  -F "video=@sample.mp4" -F "modules=m25,m26,m27,m28,m29" -F "every_n_frames=10" \
+  | python3 -m json.tool
+```
+> Rulează sincron implicit; returnează `modules_run`, `n_frames_extracted`, evidence per modul.
+> Pentru async: adaugă `-F "async_mode=true"` → primești `job_id`, apoi
+> `GET /api/status/{job_id}` și `GET /api/result/{job_id}`.
+
+## 8. Web — căutare & fact-checking evidence  (port 51100)  ✅
+
+```bash
+# cautare (SearXNG + provideri); ATENTIE: campul e "queries" (lista)
+curl -s http://$H:51100/v1/search -H 'Content-Type: application/json' -d '{
+  "queries":["deepfake detection 2024"],"max_results":5
+}' | python3 -m json.tool
+
+# fetch pagini (camp "urls" = LISTA)
+curl -s http://$H:51100/v1/fetch -H 'Content-Type: application/json' -d '{
+  "urls":["https://en.wikipedia.org/wiki/Deepfake"]
+}' | python3 -m json.tool
+
+# gather (cautare + fetch + evidence pack); camp "claim", max_search_results >= 5
+curl -s http://$H:51100/v1/gather -H 'Content-Type: application/json' -d '{
+  "claim":"Deepfakes can be detected by AI","max_search_results":5
+}' | python3 -m json.tool
+
+curl -s http://$H:51100/v1/info | python3 -m json.tool
+```
+
+## 9. Catalog — service discovery  (port 11000, doar prin gateway/intern)  ✅
+
+`catalog-api` nu e publicat pe host; se accesează prin **gateway** (Bearer) sau din rețea.
+
+```bash
+TOK=didi-local-dev-token-123
+curl -s http://$H:11000/catalog/v1/status     -H "Authorization: Bearer $TOK" | python3 -m json.tool
+curl -s http://$H:11000/catalog/v1/components -H "Authorization: Bearer $TOK" | python3 -m json.tool
+curl -s http://$H:11000/catalog/v1/models     -H "Authorization: Bearer $TOK" | python3 -m json.tool
+```
+
+## 10. Gateway — nginx reverse proxy + auth  (port 11000)  ✅
+
+```bash
+TOK=didi-local-dev-token-123
+curl -s http://$H:11000/health                                    # fara auth -> ok
+curl -s http://$H:11000/web/v1/info     -H "Authorization: Bearer $TOK"   # rutare -> web
+curl -s http://$H:11000/catalog/v1/status -H "Authorization: Bearer $TOK" # rutare -> catalog
+curl -s http://$H:11000/web/v1/info                               # fara token -> 401
+```
+
+## 11. Brain — RAG + fact-checking + cache  (port 8090)  ✅
+
+```bash
+# cautare in baza de cunostinte (contract web; camp "queries")
+curl -s http://$H:8090/v1/search -H 'Content-Type: application/json' -d '{
+  "queries":["deepfake"],"max_results":3
+}' | python3 -m json.tool
+
+# gather (RAG complet: search -> fetch -> evidence); camp "claim"
+curl -s http://$H:8090/v1/gather -H 'Content-Type: application/json' -d '{
+  "claim":"Deepfakes can be detected by AI"
+}' | python3 -m json.tool
+
+# statistici atom store (folosit de dashboard "Brain — fact status")
+curl -s http://$H:8090/v1/analysis_atom/stats/extended | python3 -m json.tool
+
+curl -s http://$H:8090/health
+```
+> `results: []` la început e normal — baza de cunoștințe e goală până se indexează conținut.
+
+---
+
+## Smoke-test rapid (toate componentele, un singur script)
+
+```bash
+export H=10.11.10.18; TOK=didi-local-dev-token-123
+ok(){ printf "%-26s %s\n" "$1" "$2"; }
+ok "llm"        "$(curl -s -m60 http://$H:14011/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"qwen3.5","messages":[{"role":"user","content":"hi"}],"max_tokens":5}' -o /dev/null -w %{http_code})"
+ok "embeddings" "$(curl -s -m20 http://$H:14100/v1/embeddings -H 'Content-Type: application/json' -d '{"model":"bge-m3","input":"x"}' -o /dev/null -w %{http_code})"
+ok "rerank"     "$(curl -s -m20 http://$H:14200/v1/rerank -H 'Content-Type: application/json' -d '{"model":"bge-reranker-v2-m3","query":"a","documents":["a","b"]}' -o /dev/null -w %{http_code})"
+ok "audio"      "$(curl -s -m10 http://$H:54300/health -o /dev/null -w %{http_code})"
+ok "video"      "$(curl -s -m10 http://$H:54600/health -o /dev/null -w %{http_code})"
+ok "extractors" "$(curl -s -m10 http://$H:54400/health -o /dev/null -w %{http_code})"
+ok "forensic"   "$(curl -s -m10 http://$H:8085/health -o /dev/null -w %{http_code})"
+ok "web"        "$(curl -s -m10 http://$H:51100/health -o /dev/null -w %{http_code})"
+ok "gateway"    "$(curl -s -m10 http://$H:11000/health -o /dev/null -w %{http_code})"
+ok "catalog"    "$(curl -s -m10 http://$H:11000/catalog/v1/status -H "Authorization: Bearer $TOK" -o /dev/null -w %{http_code})"
+ok "brain"      "$(curl -s -m10 http://$H:8090/health -o /dev/null -w %{http_code})"
+# 200 peste tot = totul up
+```
+
+---
+
+## Probleme rezolvate (istoric)
+- **sentiment / output JSON** ✅ — gateway-ul LLM injectează acum `enable_thinking:false`
+  pe vLLM (`LLM_VLLM_DISABLE_THINKING`, default true), deci modelul de reasoning întoarce
+  JSON curat pentru sentiment/OCR/semantic/brain.
+- **catalog `degraded`** ✅ — aliniat prin `CATALOG_LLM_URL/AUDIO_URL/VIDEO_URL/WEB_URL`
+  în compose → acum `healthy`.
+- **brain taxonomy 401** ✅ — creat token atomic (`atomic-server token create`), setat
+  `ATOMIC_TOKEN`, seed-uită taxonomia (79 tag-uri în atomic), parolele postgres aliniate
+  (`infra/.env`). Scheduler-ul feed-uiește acum cu `errors=0`.
+
+## Rămas minor (non-blocant)
+- **atomic embeddings** — atomic-server logează „OpenRouter API key not configured, skipping
+  embedding". Atomic-server și-ar face vectori proprii prin OpenRouter; nu e nevoie — brain
+  folosește bge-m3 local pentru embeddings. Doar un warning în logul atomic.
--- a/ai_platform/local_gpu_stack/deploy.sh
+++ b/ai_platform/local_gpu_stack/deploy.sh
@ -0,0 +1,298 @@
+#!/usr/bin/env bash
+# =============================================================================
+# DiDi LOT 1 — deploy progresiv „cărămidă cu cărămidă".
+# Pentru fiecare componentă: build/pull + download → seed (dacă e cazul) →
+# TEST funcțional → doar dacă testul TRECE corect → trece la următoarea.
+# Idempotent. STOP la primul test picat.
+#
+#   ./deploy.sh            # tot, de la preflight la dashboard
+#   ./deploy.sh --from web # sare peste cărămizile dinainte de 'web'
+#   ./deploy.sh --only llm # rulează o singură cărămidă (presupune deps up)
+# =============================================================================
+set -uo pipefail
+cd "$(dirname "$0")"
+HERE=$(pwd)
+export COMPOSE_PROJECT_NAME=local_gpu_stack
+COMPOSE="${HERE}/docker-compose.yml"
+ENV_FILE="${HERE}/.env"
+MODULES="${HERE}/../modules"
+DASH_DIR="${MODULES}/dashboard/deploy"
+GW_DIR="${MODULES}/gateway/deploy"
+BRAIN_DIR="${MODULES}/didi_brain"
+ASSETS="${HERE}/test_assets"
+
+source "${HERE}/lib.sh"
+source "${HERE}/seed/seed_catalog.sh"
+source "${HERE}/seed/seed_brain.sh"
+
+# ---- .env -------------------------------------------------------------------
+[ -f "$ENV_FILE" ] || { cp "${HERE}/.env.example" "$ENV_FILE"; info ".env creat din .env.example"; }
+set -a; source "$ENV_FILE"; set +a
+# auto-detect HOST_IP daca e gol (si scrie-l inapoi in .env pt interpolarea compose)
+if [ -z "${HOST_IP:-}" ]; then
+  HOST_IP=$(hostname -I 2>/dev/null | awk '{print $1}')
+  [ -z "$HOST_IP" ] && HOST_IP=$(ip -4 route get 1.1.1.1 2>/dev/null | awk '{for(i=1;i<=NF;i++)if($i=="src")print $(i+1)}')
+  [ -n "$HOST_IP" ] && { sed -i "s#^HOST_IP=.*#HOST_IP=${HOST_IP}#" "$ENV_FILE"; info "HOST_IP detectat automat: ${HOST_IP}"; }
+fi
+export HOST_IP
+H="${HOST_IP:-127.0.0.1}"
+# MODELS_DIR: zona locala unica pentru TOATE modelele (default: ai_platform/models)
+if [ -z "${MODELS_DIR:-}" ]; then
+  MODELS_DIR="$(cd "${HERE}/.." && pwd)/models"
+  sed -i "s#^MODELS_DIR=.*#MODELS_DIR=${MODELS_DIR}#" "$ENV_FILE"
+  info "MODELS_DIR: ${MODELS_DIR}"
+fi
+mkdir -p "$MODELS_DIR"; export MODELS_DIR
+
+# ---- arg parsing ------------------------------------------------------------
+FROM=""; ONLY=""; INSTALL_DEPS=0
+while [ $# -gt 0 ]; do case "$1" in
+  --from) FROM=$2; shift 2;; --only) ONLY=$2; shift 2;;
+  --install-deps) INSTALL_DEPS=1; shift;;
+  *) die "argument necunoscut: $1";; esac; done
+SKIP=0; [ -n "$FROM" ] && SKIP=1
+should_run(){ # $1 = brick name
+  [ -n "$ONLY" ] && { [ "$ONLY" = "$1" ]; return; }
+  [ "$SKIP" = 1 ] && { [ "$FROM" = "$1" ] && SKIP=0; }
+  [ "$SKIP" = 0 ]
+}
+
+# =============================================================================
+# PREFLIGHT  (verifică + instruiește; nu instalează)
+# =============================================================================
+# =============================================================================
+# INSTALL DEPS (optional, --install-deps; necesita sudo; Ubuntu/Debian)
+# =============================================================================
+install_deps() {
+  brick "INSTALL DEPS — Docker + NVIDIA Container Toolkit (sudo)"
+  command -v sudo >/dev/null || die "sudo necesar pentru --install-deps"
+  if ! command -v docker >/dev/null; then
+    info "instalez Docker Engine..."; curl -fsSL https://get.docker.com | sudo sh || die "instalare Docker a picat"
+    sudo usermod -aG docker "$USER" || true
+    warn "Adăugat în grupul docker — poate fi nevoie de re-login pentru ca 'docker' să meargă fără sudo."
+  else ok "Docker prezent"; fi
+  if ! docker info 2>/dev/null | grep -qiE 'nvidia|cdi' && ! sudo docker info 2>/dev/null | grep -qiE 'nvidia|cdi'; then
+    info "instalez nvidia-container-toolkit..."
+    curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
+    curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
+      | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
+      | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list >/dev/null
+    sudo apt-get update -y && sudo apt-get install -y nvidia-container-toolkit || die "instalare toolkit a picat"
+    sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker
+    ok "nvidia-container-toolkit instalat & Docker reconfigurat"
+  else ok "NVIDIA runtime prezent"; fi
+  command -v nvidia-smi >/dev/null || warn "Driver NVIDIA lipsește — instalează manual (ex: sudo ubuntu-drivers install) și REBOOT, apoi re-rulează deploy.sh."
+}
+
+preflight() {
+  brick "PREFLIGHT — unelte & GPU"
+  command -v docker >/dev/null || die "Docker lipsește. Instalează: https://docs.docker.com/engine/install/ubuntu/"
+  if ! docker ps >/dev/null 2>&1; then
+    [ "$USE_SG" = 1 ] && warn "rulez docker prin 'sg docker' (userul nu e în grupul docker activ)" \
+      || die "Nu pot accesa Docker. Adaugă userul în grup:  sudo usermod -aG docker \$USER  (apoi re-login)"
+  fi
+  docker compose version >/dev/null 2>&1 || die "Docker Compose v2 lipsește (plugin 'docker compose')."
+  command -v nvidia-smi >/dev/null || die "nvidia-smi lipsește — instalează driverul NVIDIA (>=535) + CUDA."
+  local ngpu; ngpu=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | wc -l)
+  [ "$ngpu" -ge 1 ] || die "Niciun GPU detectat de nvidia-smi."
+  if ! docker info 2>/dev/null | grep -qiE 'nvidia|cdi'; then
+    die "Runtime-ul NVIDIA pt Docker lipsește. Instalează nvidia-container-toolkit:
+       https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html
+       apoi:  sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker"
+  fi
+  command -v curl >/dev/null || die "curl lipsește (apt-get install -y curl)."
+  command -v python3 >/dev/null || die "python3 lipsește."
+  local freeg; freeg=$(df -BG --output=avail / | tail -1 | tr -dc '0-9')
+  [ "${freeg:-0}" -ge 80 ] || warn "Doar ${freeg}G liberi pe / — Qwen are ~72GB. Recomandat >100G."
+  ok "Docker $(docker --version | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -1) · ${ngpu} GPU · ${freeg}G liberi"
+}
+
+# =============================================================================
+# NETWORK
+# =============================================================================
+ensure_network() {
+  brick "REȚEA — didi-network (una singură, partajată)"
+  if docker network inspect didi-network >/dev/null 2>&1; then ok "didi-network există"
+  else docker network create didi-network >/dev/null && ok "didi-network creată"; fi
+}
+
+# =============================================================================
+# BUILD images (app) + PULL (vLLM/searxng)
+# =============================================================================
+build_images() {
+  brick "BUILD — imagini aplicație"
+  local m
+  declare -A MAP=(
+    [llm-inference]=didiai-llm-inference [embeddings]=didiai-embeddings [rerank]=didiai-rerank
+    [audio]=didiai-audio [video-analysis]=didiai-video-analysis [extractors]=didiai-extractors
+    [web]=didiai-web [catalog-api]=didiai-catalog-api )
+  for m in "${!MAP[@]}"; do
+    printf '  build %-16s ' "$m"
+    if docker build -q -t "${MAP[$m]}:audit" -f "${MODULES}/$m/deploy/Dockerfile" "${MODULES}/$m" >/tmp/build_$m.log 2>&1; then ok "ok"
+    else err "FAIL"; tail -15 /tmp/build_$m.log | sed 's/^/      /'; die "build $m a picat"; fi
+  done
+  # forensic (layout non-standard: Dockerfile la root)
+  printf '  build %-16s ' "forensic"
+  if docker build -q -t didiai-forensic:audit -f "${MODULES}/forensic_features/Dockerfile" "${MODULES}/forensic_features" >/tmp/build_forensic.log 2>&1; then ok "ok"
+  else err "FAIL"; tail -15 /tmp/build_forensic.log | sed 's/^/      /'; die "build forensic a picat"; fi
+  gate "imagini prezente" bash -c 'docker images --format "{{.Repository}}" | grep -q didiai-llm-inference'
+}
+
+pull_images() {
+  brick "PULL — runtime-uri vLLM + SearXNG"
+  local img
+  for img in vllm/vllm-openai:v0.8.5 vllm/vllm-openai:qwen3_5 docker.io/searxng/searxng:latest; do
+    printf '  pull %-34s ' "$img"
+    if docker image inspect "$img" >/dev/null 2>&1; then ok "deja prezent"
+    elif docker pull -q "$img" >/dev/null 2>&1; then ok "ok"
+    else die "pull $img a picat (verifică rețeaua / Docker Hub)"; fi
+  done
+}
+
+# =============================================================================
+# PRE-DOWNLOAD modele (cu progres)
+# =============================================================================
+predownload_models() {
+  brick "DOWNLOAD — modele în MODELS_DIR (offline-first, un folder per model)"
+  mkdir -p "${MODELS_DIR}"
+  # mini-script de download (evită iadul de ghilimele prin sg docker)
+  local dl="${HERE}/.dl_model.py"
+  printf 'import os\nfrom huggingface_hub import snapshot_download\nsnapshot_download(os.environ["REPO"])\nprint("OK", os.environ["REPO"])\n' > "$dl"
+  # repo HF -> subfolder cache (models--<org>--<name>)
+  local repo cache
+  for repo in "$BUSTER_MODEL" "$LLM_MODEL" "$EMB_MODEL" "$RERANK_MODEL" "$WHISPER_REPO" "$NER_MODEL" "$NER_BACKBONE"; do
+    cache="models--$(printf '%s' "$repo" | sed 's#/#--#')"
+    if [ -d "${MODELS_DIR}/${cache}/snapshots" ]; then
+      ok "prezent local: ${repo}"
+    else
+      info "download: ${repo}"
+      docker run --rm -e HF_HOME=/c -e "REPO=${repo}" \
+        -v "${MODELS_DIR}:/c/hub" -v "${dl}:/dl.py" \
+        --entrypoint python3 vllm/vllm-openai:qwen3_5 /dl.py \
+        || die "download ${repo} a picat."
+      ok "${repo} descărcat"
+    fi
+  done
+  rm -f "$dl"
+  ok "Toate modelele sunt în ${MODELS_DIR}"
+}
+
+# =============================================================================
+# TEST helpers (output corect, nu doar 200)
+# =============================================================================
+t_embeddings(){ post_json "http://$H:14100/v1/embeddings" '{"model":"bge-m3","input":"test"}' 20 | python3 -c 'import sys,json;v=json.load(sys.stdin)["data"][0]["embedding"];assert len(v)==1024;print(f"dim={len(v)}")'; }
+t_rerank(){ post_json "http://$H:14200/v1/rerank" '{"model":"bge-reranker-v2-m3","query":"deepfake detection","documents":["o pisica","sistem detectie deepfake video"]}' 20 | python3 -c 'import sys,json;r=json.load(sys.stdin)["results"];assert r[0]["index"]==1;print("top_idx=%d score=%.3f"%(r[0]["index"],r[0]["relevance_score"]))'; }
+t_llm(){ post_json "http://$H:14011/v1/chat/completions" '{"model":"qwen3.5","messages":[{"role":"user","content":"Spune un singur cuvant: salut"}],"max_tokens":20}' 90 | python3 -c 'import sys,json;c=json.load(sys.stdin)["choices"][0]["message"]["content"].strip();assert c and not c.lower().startswith("thinking");print(repr(c[:40]))'; }
+t_audio(){ curl -s -m120 "http://$H:54300/v1/audio/transcriptions" -F "file=@${ASSETS}/audio.wav" -F "model=${WHISPER_MODEL}" | python3 -c 'import sys,json;t=json.load(sys.stdin)["text"].strip();assert len(t)>0;print(f"text[:40]={t[:40]!r}")'; }
+t_video(){ curl -s -m180 "http://$H:54600/analyze/video" -F "file=@${ASSETS}/sample.mp4" | python3 -c 'import sys,json;d=json.load(sys.stdin);assert d["verdict"] in ("REAL","FAKE","UNCERTAIN");print("verdict=%s frames=%s"%(d["verdict"],d.get("frames_analyzed")))'; }
+t_ner(){ post_json "http://$H:54400/v1/ner" '{"text":"Klaus Iohannis s-a intalnit cu Emmanuel Macron la Bucuresti."}' 60 | python3 -c 'import sys,json;e=json.load(sys.stdin)["results"]["entities"];assert len(e)>=2;print(f"entities={len(e)}")'; }
+t_detect(){ curl -s -m90 "http://$H:54400/v1/detect" -F "file=@${ASSETS}/frame.jpg" | python3 -c 'import sys,json;d=json.load(sys.stdin)["results"]["detections"];assert len(d)>=1;print(f"objects={len(d)}")'; }
+t_metadata(){ curl -s -m40 "http://$H:54400/v1/metadata" -F "file=@${ASSETS}/sample.mp4" | python3 -c 'import sys,json;d=json.load(sys.stdin);assert d["analyses"]["integrity"]["ok"];print("integrity ok")'; }
+t_sentiment(){ post_json "http://$H:54400/v1/sentiment" '{"text":"Produsul este excelent, sunt foarte multumit!"}' 60 | python3 -c 'import sys,json;d=json.load(sys.stdin);assert d["ok"];print("label=%s"%d["results"].get("label"))'; }
+t_forensic(){ curl -s -m15 "http://$H:8085/api/forensic-modules" | python3 -c 'import sys,json;m=json.load(sys.stdin)["available_modules"];assert len(m)>=3;print(f"modules={len(m)}")'; }
+t_web(){ post_json "http://$H:51100/v1/search" '{"queries":["deepfake detection"],"max_results":3}' 40 | python3 -c 'import sys,json;r=json.load(sys.stdin)["results"];assert len(r)>0;print(f"results={len(r)}")'; }
+t_catalog(){ curl -s -m15 "http://$H:11000/catalog/v1/status" -H "Authorization: Bearer ${GATEWAY_API_TOKEN}" | python3 -c 'import sys,json;d=json.load(sys.stdin);assert d["status"]=="healthy";print("catalog healthy")'; }
+t_gateway(){ local a b; a=$(curl -s -m10 -o /dev/null -w '%{http_code}' "http://$H:11000/health"); b=$(curl -s -m10 -o /dev/null -w '%{http_code}' "http://$H:11000/web/v1/info"); [ "$a" = 200 ] && [ "$b" = 401 ] && echo "health=200 noauth=401" || { echo "health=$a noauth=$b (astept 200/401)"; return 1; }; }
+t_brain(){ curl -s -m15 "http://$H:8090/v1/analysis_atom/stats/extended" | python3 -c 'import sys,json;json.load(sys.stdin);print("stats ok")' && post_json "http://$H:8090/v1/search" '{"queries":["deepfake"],"max_results":2}' 30 | python3 -c 'import sys,json;assert "results" in json.load(sys.stdin);print("search ok")'; }
+t_dashboard(){ curl -s -m15 "http://$H:51300/api/monitoring/services" | python3 -c 'import sys,json;d=json.load(sys.stdin);s=d if isinstance(d,list) else d.get("services",[]);h=sum(x.get("status")=="healthy" for x in s);assert h==len(s) and h>=11,f"{h}/{len(s)}";print(f"{h}/{len(s)} healthy")'; }
+
+prepare_assets() {
+  mkdir -p "$ASSETS"
+  [ -f "${ASSETS}/sample.mp4" ] || cp "$(find ${HERE}/../../artefacte_lot1/runs_deepfake -iname '*.mp4' -printf '%s\t%p\n' | sort -n | head -1 | cut -f2)" "${ASSETS}/sample.mp4"
+  if [ ! -f "${ASSETS}/frame.jpg" ] || [ ! -f "${ASSETS}/audio.wav" ]; then
+    docker cp "${ASSETS}/sample.mp4" didiAI-extractors:/tmp/s.mp4 >/dev/null
+    docker exec didiAI-extractors sh -c 'ffmpeg -y -i /tmp/s.mp4 -frames:v 1 /tmp/frame.jpg -vn -ar 16000 -ac 1 -t 8 /tmp/audio.wav' >/dev/null 2>&1
+    docker cp didiAI-extractors:/tmp/frame.jpg "${ASSETS}/frame.jpg" >/dev/null
+    docker cp didiAI-extractors:/tmp/audio.wav "${ASSETS}/audio.wav" >/dev/null
+  fi
+}
+
+# =============================================================================
+# CĂRĂMIZI (fiecare: up dependențe → health → test → gate)
+# =============================================================================
+brick_embeddings(){ brick "EMBEDDINGS (bge-m3)"; dc up -d vllm-embeddings embeddings-api
+  wait_health "http://$H:14100/health" 180 "embeddings-api" || die "embeddings-api nu pornește"; gate "embeddings dim 1024" t_embeddings; }
+brick_rerank(){ brick "RERANK (bge-reranker-v2-m3)"; dc up -d vllm-rerank rerank-api
+  wait_health "http://$H:14200/health" 180 "rerank-api" || die "rerank-api nu pornește"; gate "rerank ranking" t_rerank; }
+brick_llm(){ brick "LLM (Qwen3.5-35B-A3B)"; dc up -d vllm-qwen llm-api
+  wait_health "http://$H:14001/v1/models" 900 "vLLM Qwen (load+compile)" || die "vLLM Qwen nu pornește"
+  wait_health "http://$H:14011/health" 60 "llm-api" || die "llm-api nu pornește"; gate "LLM output curat" t_llm; }
+brick_extractors(){ brick "EXTRACTORS (NER/YOLO/metadata/OCR/sentiment)"; dc up -d extractors
+  wait_health "http://$H:54400/health" 120 "extractors" || die "extractors nu pornește"
+  prepare_assets
+  gate "metadata" t_metadata; gate "NER (GLiNER)" t_ner; gate "detect (YOLO)" t_detect; gate "sentiment (LLM)" t_sentiment; }
+brick_audio(){ brick "AUDIO (Whisper large-v3-turbo)"; dc up -d audio-api
+  wait_health "http://$H:54300/health" 240 "audio (download model)" || die "audio nu pornește"; gate "transcriere" t_audio; }
+brick_video(){ brick "VIDEO / BusterX++ (deepfake)"; dc up -d vllm-buster video-api
+  wait_health "http://$H:54500/health" 300 "vLLM BusterX" || die "vLLM BusterX nu pornește"
+  wait_health "http://$H:54600/health" 60 "video-api" || die "video-api nu pornește"; gate "verdict deepfake" t_video; }
+brick_forensic(){ brick "FORENSIC (rPPG/lip-sync/forgery/lighting)"; dc up -d forensic
+  wait_health "http://$H:8085/health" 120 "forensic" || die "forensic nu pornește"; gate "module forensice" t_forensic; }
+brick_web(){ brick "WEB + SearXNG (căutare)"
+  # curăță o instanță searxng pornită manual (docker run) ca s-o preia compose
+  if docker inspect didiAI-web-searxng >/dev/null 2>&1 && \
+     [ -z "$(docker inspect didiAI-web-searxng --format '{{ index .Config.Labels "com.docker.compose.project" }}' 2>/dev/null)" ]; then
+    docker rm -f didiAI-web-searxng >/dev/null 2>&1 || true
+  fi
+  dc up -d searxng web-api
+  wait_health "http://$H:51100/health" 120 "web-api" || die "web-api nu pornește"
+  sleep 8; gate "căutare web" t_web; }
+brick_catalog(){ brick "CATALOG (service discovery)"; dc up -d catalog-api
+  sleep 6; }   # health-ul agregat depinde de gateway; testat după gateway
+brick_gateway(){ brick "GATEWAY (nginx + auth)"
+  ( cd "$GW_DIR"; grep -q . .env 2>/dev/null || echo "GATEWAY_API_TOKEN=${GATEWAY_API_TOKEN}" > .env; docker compose up -d )
+  wait_health "http://$H:11000/health" 60 "gateway" || die "gateway nu pornește"
+  gate "gateway auth+rutare" t_gateway; gate "catalog agregat" t_catalog; }
+brick_brain(){ brick "BRAIN (RAG + fact-check + cache)"
+  # render env din master, up postgres+atomic, aliniază parola, token, taxonomie, up api+scheduler
+  brain_write_env "$POSTGRES_PASSWORD" "${ATOMIC_TOKEN:-PLACEHOLDER}"
+  ( cd "$BRAIN_DIR"; docker compose -f infra/docker-compose.yml up -d postgres atomic-server )
+  wait_health "http://$H:8088/health" 120 "atomic-server" || warn "atomic health întârzie"
+  brain_align_pg "$POSTGRES_PASSWORD"
+  local tok; tok=$(brain_ensure_token "$POSTGRES_PASSWORD" "${ATOMIC_TOKEN:-}")
+  [ -n "$tok" ] || die "nu am putut crea tokenul atomic"
+  if [ "$tok" != "${ATOMIC_TOKEN:-}" ]; then
+    sed -i "s#^ATOMIC_TOKEN=.*#ATOMIC_TOKEN=${tok}#" "$ENV_FILE"; ATOMIC_TOKEN="$tok"; ok "token atomic creat & salvat în .env"
+  fi
+  brain_write_env "$POSTGRES_PASSWORD" "$tok"
+  ( cd "$BRAIN_DIR"; docker compose -f infra/docker-compose.yml up -d brain-api scheduler )
+  wait_health "http://$H:8090/health" 120 "brain-api" || die "brain-api nu pornește"
+  brain_seed_taxonomy
+  ( cd "$BRAIN_DIR"; docker compose -f infra/docker-compose.yml restart brain-api scheduler ) >/dev/null
+  wait_health "http://$H:8090/health" 60 "brain-api (post-seed)" || die "brain-api nu revine"
+  gate "brain stats+search" t_brain; }
+brick_dashboard(){ brick "DASHBOARD (UI + monitorizare)"
+  ( cd "$DASH_DIR"
+    cat > .env <<EOF
+DASHBOARD_DB_USER=${DASHBOARD_DB_USER}
+DASHBOARD_DB_PASSWORD=${DASHBOARD_DB_PASSWORD}
+DASHBOARD_DB_NAME=${DASHBOARD_DB_NAME}
+DASHBOARD_EXTERNAL_URL=http://${H}:51300
+DASHBOARD_STAGING_MODE=${DASHBOARD_STAGING_MODE}
+VITE_STAGING_MODE=${DASHBOARD_STAGING_MODE}
+DASHBOARD_LLM_API_URL=http://didiAI-llm-api:14011
+DASHBOARD_EMBEDDINGS_HEALTH_URL=http://didiAI-embeddings-api:14100/health
+DASHBOARD_RERANK_HEALTH_URL=http://didiAI-rerank-api:14200/health
+DASHBOARD_AUDIO_HEALTH_URL=http://didiAI-audio:54300/v1/info
+DASHBOARD_VIDEO_HEALTH_URL=http://didiAI-video-api:54600/v1/info
+DASHBOARD_EXTRACTORS_HEALTH_URL=http://didiAI-extractors:54400/health
+DASHBOARD_FORENSIC_HEALTH_URL=http://didiAI-forensic:8080/health
+DASHBOARD_WEB_HEALTH_URL=http://didiAI-web-api:51100/v1/info
+EOF
+    docker compose --profile dashboard up -d --build )
+  wait_health "http://$H:51300/health" 120 "dashboard" || die "dashboard nu pornește"
+  seed_catalog
+  sleep 5; gate "dashboard 11/11 healthy" t_dashboard; }
+
+# =============================================================================
+# RUN
+# =============================================================================
+[ "$INSTALL_DEPS" = 1 ] && install_deps
+if [ -z "$ONLY" ]; then preflight; ensure_network; build_images; pull_images; predownload_models; fi
+for b in embeddings rerank llm extractors audio video forensic web catalog gateway brain dashboard; do
+  should_run "$b" && "brick_$b"
+done
+
+brick "GATA"
+ok "Toate cărămizile au trecut testele. Dashboard: http://${H}:51300/admin-ai/"
+echo "Smoke-test complet:  vezi ${HERE}/TESTING.md"
--- a/ai_platform/local_gpu_stack/docker-compose.yml
+++ b/ai_platform/local_gpu_stack/docker-compose.yml
@ -0,0 +1,310 @@
+# Local GPU inference stack — 10.11.10.18 (2x H200 NVL)
+# GPU 0: Qwen3.5-35B-A3B (LLM) + Whisper (audio)
+# GPU 1: BusterX++ (deepfake video) + bge-m3 (embeddings) + bge-reranker-v2-m3 (rerank)
+# All services on the shared `didi-network` so the dashboard can probe them by name.
+
+networks:
+  didi-network:
+    external: true
+
+services:
+  # ===================== vLLM model servers =====================
+
+  # BusterX++ — served OFFLINE from the on-disk HF cache (no download)
+  vllm-buster:
+    container_name: didiAI-video-vllm-buster
+    # qwen3_5 image (vLLM 0.16 / transformers 4.57) — BusterX++ config was saved
+    # with transformers 5.x and the older 0.8.5 image can't parse its nested text_config.
+    image: vllm/vllm-openai:qwen3_5
+    networks: [didi-network]
+    ports: ["54500:54500"]
+    volumes:
+      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
+    environment:
+      - HF_HOME=/root/.cache/huggingface
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+    command: >
+      --model l8cv/BusterX_plusplus
+      --host 0.0.0.0 --port 54500
+      --served-model-name busterx
+      --tensor-parallel-size 1
+      --max-model-len 32768
+      --gpu-memory-utilization 0.25
+      --trust-remote-code
+      --enable-prefix-caching
+      --disable-log-requests
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ['1']
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  # Qwen3.5-35B-A3B — downloads ~72GB on first start
+  vllm-qwen:
+    container_name: didiAI-vllm-qwen3.5
+    image: vllm/vllm-openai:qwen3_5
+    networks: [didi-network]
+    ports: ["14001:14001"]
+    volumes:
+      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
+    environment:
+      - HF_HOME=/root/.cache/huggingface
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+      - VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+    command: >
+      --model Qwen/Qwen3.5-35B-A3B
+      --host 0.0.0.0 --port 14001
+      --served-model-name qwen3.5
+      --tensor-parallel-size 1
+      --max-model-len 32000
+      --gpu-memory-utilization 0.65
+      --trust-remote-code
+      --disable-log-requests
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ['0']
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  # bge-m3 embeddings — downloads ~2GB on first start
+  vllm-embeddings:
+    container_name: didiAI-vllm-embeddings
+    image: vllm/vllm-openai:v0.8.5
+    networks: [didi-network]
+    ports: ["14101:14101"]
+    volumes:
+      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
+    environment:
+      - HF_HOME=/root/.cache/huggingface
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+    command: >
+      --model BAAI/bge-m3
+      --host 0.0.0.0 --port 14101
+      --served-model-name bge-m3
+      --task embed
+      --gpu-memory-utilization 0.10
+      --trust-remote-code
+      --disable-log-requests
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ['1']
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  # bge-reranker-v2-m3 — downloads ~2GB on first start
+  vllm-rerank:
+    container_name: didiAI-vllm-rerank
+    image: vllm/vllm-openai:v0.8.5
+    networks: [didi-network]
+    ports: ["14201:14201"]
+    volumes:
+      - ${MODELS_DIR}:/root/.cache/huggingface/hub:ro
+    environment:
+      - HF_HOME=/root/.cache/huggingface
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+    command: >
+      --model BAAI/bge-reranker-v2-m3
+      --host 0.0.0.0 --port 14201
+      --served-model-name bge-reranker-v2-m3
+      --task score
+      --gpu-memory-utilization 0.10
+      --trust-remote-code
+      --disable-log-requests
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ['1']
+              capabilities: [gpu]
+    restart: unless-stopped
+
+  # ===================== FastAPI app wrappers =====================
+
+  llm-api:
+    container_name: didiAI-llm-api
+    image: didiai-llm-inference:audit
+    networks: [didi-network]
+    ports: ["14011:14011"]
+    environment:
+      - LLM_HOST=0.0.0.0
+      - LLM_PORT=14011
+      - LLM_DEFAULT_BACKEND=vllm
+      - LLM_ENABLE_VLLM=true
+      - LLM_ENABLE_LLAMACPP=false
+      - LLM_VLLM_BASE_URL=http://didiAI-vllm-qwen3.5:14001
+      - LLM_DEFAULT_MODEL=qwen3.5
+      - LLM_EXTERNAL_URL=http://${HOST_IP}:14011
+    restart: unless-stopped
+
+  embeddings-api:
+    container_name: didiAI-embeddings-api
+    image: didiai-embeddings:audit
+    networks: [didi-network]
+    ports: ["14100:14100"]
+    environment:
+      - EMB_HOST=0.0.0.0
+      - EMB_PORT=14100
+      - EMB_DEFAULT_BACKEND=vllm
+      - EMB_ENABLE_VLLM=true
+      - EMB_ENABLE_LLAMACPP=false
+      - EMB_VLLM_BASE_URL=http://didiAI-vllm-embeddings:14101
+      - EMB_VLLM_MODEL=bge-m3
+      - EMB_EXTERNAL_URL=http://${HOST_IP}:14100
+    restart: unless-stopped
+
+  rerank-api:
+    container_name: didiAI-rerank-api
+    image: didiai-rerank:audit
+    networks: [didi-network]
+    ports: ["14200:14200"]
+    environment:
+      - RERANK_HOST=0.0.0.0
+      - RERANK_PORT=14200
+      - RERANK_DEFAULT_BACKEND=vllm
+      - RERANK_ENABLE_VLLM=true
+      - RERANK_ENABLE_LLAMACPP=false
+      - RERANK_VLLM_BASE_URL=http://didiAI-vllm-rerank:14201
+      - RERANK_VLLM_MODEL=bge-reranker-v2-m3
+      - RERANK_EXTERNAL_URL=http://${HOST_IP}:14200
+    restart: unless-stopped
+
+  video-api:
+    container_name: didiAI-video-api
+    image: didiai-video-analysis:audit
+    networks: [didi-network]
+    ports: ["54600:54600"]
+    volumes:
+      - ./runs:/app/runs
+    environment:
+      - VIDEO_ANALYSIS_HOST=0.0.0.0
+      - VIDEO_ANALYSIS_PORT=54600
+      - VIDEO_ANALYSIS_EXTERNAL_URL=http://${HOST_IP}:54600
+      - VIDEO_ANALYSIS_VLLM_BASE_URL=http://didiAI-video-vllm-buster:54500
+      - VIDEO_ANALYSIS_VLLM_MODEL=busterx
+      - VIDEO_ANALYSIS_RUNS_DIR=/app/runs
+      - VIDEO_ANALYSIS_SEMANTIC_LLM_BASE_URL=http://didiAI-llm-api:14011
+      - VIDEO_ANALYSIS_SEMANTIC_AGGREGATION_MODEL=qwen3.5
+      - VIDEO_ANALYSIS_SEMANTIC_LLM_API_KEY=none
+    restart: unless-stopped
+
+  # ===================== Non-GPU services =====================
+
+  catalog-api:
+    container_name: didiAI-catalog-api
+    image: didiai-catalog-api:audit
+    networks: [didi-network]
+    environment:
+      - CATALOG_HOST=0.0.0.0
+      - CATALOG_PORT=11000
+      - CATALOG_EXTERNAL_URL=http://${HOST_IP}
+      - CATALOG_LLM_URL=http://didiAI-llm-api:14011
+      - CATALOG_AUDIO_URL=http://didiAI-audio:54300
+      - CATALOG_VIDEO_URL=http://didiAI-video-api:54600
+      - CATALOG_WEB_URL=http://didiAI-web-api:51100
+    restart: unless-stopped
+
+  extractors:
+    container_name: didiAI-extractors
+    image: didiai-extractors:audit
+    networks: [didi-network]
+    ports: ["54400:54400"]
+    volumes:
+      # GLiNER (+ backbone mdeberta) servit offline din zona locala de modele
+      - ${MODELS_DIR}:/models-hf/hub:ro
+    environment:
+      - EXTRACTORS_HOST=0.0.0.0
+      - EXTRACTORS_PORT=54400
+      - EXTRACTORS_LLM_GATEWAY_URL=http://didiAI-llm-api:14011
+      - EXTRACTORS_DETECT_MODEL=/tmp/yolov8n.pt
+      - YOLO_CONFIG_DIR=/tmp/ultralytics
+      - HF_HOME=/models-hf
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+    restart: unless-stopped
+
+  forensic:
+    container_name: didiAI-forensic
+    image: didiai-forensic:audit
+    networks: [didi-network]
+    ports: ["8085:8080"]
+    environment:
+      - API_PORT=8080
+    restart: unless-stopped
+
+  # cloak — scraper SERP stealth (tier-3 fallback pentru web)
+  cloak:
+    container_name: didiAI-cloak
+    image: didiai-cloak:audit
+    networks: [didi-network]
+    ports: ["8770:8770"]
+    environment:
+      - CLOAK_HOST=0.0.0.0
+      - CLOAK_PORT=8770
+    restart: unless-stopped
+
+  # SearXNG metasearch (motorul de cautare liber folosit de web-api)
+  searxng:
+    container_name: didiAI-web-searxng
+    image: docker.io/searxng/searxng:latest
+    networks: [didi-network]
+    volumes:
+      - ../../modules/web/deploy/metasearch/searxng:/etc/searxng:ro
+    environment:
+      - SEARXNG_BASE_URL=http://${HOST_IP}:8080/
+    restart: unless-stopped
+
+  web-api:
+    container_name: didiAI-web-api
+    image: didiai-web:audit
+    networks: [didi-network]
+    ports: ["51100:51100"]
+    depends_on: [searxng]
+    environment:
+      - WEB_HOST=0.0.0.0
+      - WEB_PORT=51100
+      - WEB_EXTERNAL_URL=http://${HOST_IP}:51100
+      - WEB_SEARXNG_BASE_URL=http://didiAI-web-searxng:8080
+      - WEB_LLM_BASE_URL=http://didiAI-llm-api:14011
+    restart: unless-stopped
+
+  # Whisper STT — faster-whisper loads in-process on GPU 0
+  audio-api:
+    container_name: didiAI-audio
+    image: didiai-audio:audit
+    networks: [didi-network]
+    ports: ["54300:54300"]
+    volumes:
+      # faster-whisper foloseste download_root=cache_dir -> models-- direct sub el (nu /hub)
+      - ${MODELS_DIR}:/root/.cache/huggingface:ro
+    environment:
+      - AUDIO_HOST=0.0.0.0
+      - AUDIO_PORT=54300
+      - AUDIO_MODEL=large-v3-turbo
+      - AUDIO_DEVICE=cuda
+      - AUDIO_COMPUTE_TYPE=float16
+      - AUDIO_CACHE_DIR=/root/.cache/huggingface
+      - HF_HUB_OFFLINE=1
+      - TRANSFORMERS_OFFLINE=1
+      - AUDIO_EXTERNAL_URL=http://${HOST_IP}:54300
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids: ['0']
+              capabilities: [gpu]
+    restart: unless-stopped
--- a/ai_platform/local_gpu_stack/lib.sh
+++ b/ai_platform/local_gpu_stack/lib.sh
@ -0,0 +1,74 @@
+#!/usr/bin/env bash
+# Shared helpers for the brick-by-brick deploy. Sourced by deploy.sh.
+# Logging, test gates, health waits, and a docker shim that works whether or not
+# the current shell is in the `docker` group.
+
+set -uo pipefail
+
+# ---- colors ----------------------------------------------------------------
+if [ -t 1 ]; then
+  C_R=$'\e[31m'; C_G=$'\e[32m'; C_Y=$'\e[33m'; C_B=$'\e[36m'; C_D=$'\e[2m'; C_0=$'\e[0m'
+else
+  C_R=; C_G=; C_Y=; C_B=; C_D=; C_0=
+fi
+log()  { printf '%s\n' "$*"; }
+info() { printf '%s\n' "${C_B}»${C_0} $*"; }
+ok()   { printf '%s\n' "${C_G}✓${C_0} $*"; }
+warn() { printf '%s\n' "${C_Y}!${C_0} $*"; }
+err()  { printf '%s\n' "${C_R}✗${C_0} $*" >&2; }
+brick(){ printf '\n%s\n' "${C_B}━━━ CĂRĂMIDĂ: $* ━━━${C_0}"; }
+die()  { err "$*"; err "Deploy oprit (gate). Repară și re-rulează — e idempotent."; exit 1; }
+
+# ---- docker shim (portable: plain docker, else `sg docker`) -----------------
+USE_SG=0
+if ! docker ps >/dev/null 2>&1; then
+  if command -v sg >/dev/null 2>&1 && sg docker -c 'docker ps' >/dev/null 2>&1; then
+    USE_SG=1
+  fi
+fi
+docker() {
+  if [ "$USE_SG" = 1 ]; then
+    local a args=""
+    for a in "$@"; do args+=" $(printf '%q' "$a")"; done
+    sg docker -c "docker$args"
+  else
+    command docker "$@"
+  fi
+}
+
+# ---- compose wrapper -------------------------------------------------------
+# COMPOSE is exported by deploy.sh
+dc() { docker compose -f "$COMPOSE" "$@"; }
+
+# ---- health / test helpers -------------------------------------------------
+# wait_health <url> <timeout_s> <label>
+wait_health() {
+  local url=$1 timeout=${2:-120} label=${3:-$1} i=0
+  printf '  %s' "aștept $label "
+  while [ "$i" -lt "$timeout" ]; do
+    if [ "$(curl -s -m5 -o /dev/null -w '%{http_code}' "$url" 2>/dev/null)" = "200" ]; then
+      printf ' %sup%s\n' "$C_G" "$C_0"; return 0
+    fi
+    printf '.'; sleep 3; i=$((i+3))
+  done
+  printf '\n'; return 1
+}
+
+# gate <label> <test-command...>
+# Runs the test; on non-zero exit, STOPS the whole deploy.
+gate() {
+  local label=$1; shift
+  printf '  test: %-22s ' "$label"
+  local out rc
+  out=$("$@" 2>&1); rc=$?
+  if [ "$rc" -eq 0 ]; then
+    ok "PASS  ${C_D}${out}${C_0}"
+    return 0
+  fi
+  err "FAIL"
+  printf '%s\n' "$out" | sed 's/^/      /'
+  die "Testul '$label' nu a trecut."
+}
+
+# json_field <url> <python-expr-on-d>  -> prints value, exit 1 on error
+post_json() { curl -s -m"${4:-60}" "$1" -H 'Content-Type: application/json' -d "$2"; }
--- a/ai_platform/local_gpu_stack/reports/latest.json
+++ b/ai_platform/local_gpu_stack/reports/latest.json
@ -0,0 +1,422 @@
+{
+  "generated_at": "2026-06-27T14:46:37.623279+03:00",
+  "host": "10.11.10.18",
+  "summary": {
+    "total": 30,
+    "passed": 30,
+    "failed": 0
+  },
+  "tests": [
+    {
+      "id": "llm",
+      "name": "LLM (Qwen3.5-35B-A3B)",
+      "endpoint": "POST /v1/chat/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 342.2,
+      "result": {
+        "reply": "Un deepfake este o media manipulat, realizat prin tehnologia inteligenței artificiale, care înlocuiează cu realism înspăimântător fața sau vocea unei persoane cu cea a altcuiva."
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "llm_completions",
+      "name": "LLM · completions",
+      "endpoint": "POST /v1/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 111.8,
+      "result": {
+        "text": " un oras care atrage foarte multi turisti si"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "llm_info",
+      "name": "LLM · info",
+      "endpoint": "GET /v1/info",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 14.5,
+      "result": {
+        "keys": [
+          "resource",
+          "models",
+          "functions"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "embeddings",
+      "name": "Embeddings (bge-m3)",
+      "endpoint": "POST /v1/embeddings",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 43.8,
+      "result": {
+        "dim": 1024
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "rerank",
+      "name": "Rerank (bge-reranker-v2-m3)",
+      "endpoint": "POST /v1/rerank",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 39.3,
+      "result": {
+        "top_index": 1,
+        "score": 0.626
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "audio",
+      "name": "Audio (Whisper large-v3-turbo)",
+      "endpoint": "POST /v1/audio/transcriptions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 371.3,
+      "result": {
+        "text": "Nici nu mai știi cine este de fapt agresorul și agresatul. De ce? Simplu. Pentru că în acest moment, Uniunea Europeană...",
+        "language": "ro"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "video",
+      "name": "Video/BusterX (deepfake)",
+      "endpoint": "POST /analyze/video",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1389.8,
+      "result": {
+        "verdict": "REAL",
+        "frames_analyzed": 16
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "video_semantic",
+      "name": "Video · semantic",
+      "endpoint": "POST /analyze/video/semantic",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 18390.9,
+      "result": {
+        "num_chunks": 5,
+        "summary": "The video presents a continuous, polemical monologue delivered by a male speaker in a studio setting, likely a podcast o"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_metadata",
+      "name": "Extractors · metadata",
+      "endpoint": "POST /v1/metadata",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 284.1,
+      "result": {
+        "media_type": "video",
+        "sha256": "27d6c00786e63905"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_ner",
+      "name": "Extractors · NER (GLiNER)",
+      "endpoint": "POST /v1/ner",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 262.9,
+      "result": {
+        "entities": [
+          "Klaus Iohannis:persoană",
+          "Emmanuel Macron:persoană",
+          "Bucuresti:locație"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_detect",
+      "name": "Extractors · detect (YOLOv8)",
+      "endpoint": "POST /v1/detect",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 77.1,
+      "result": {
+        "objects": 4,
+        "labels": [
+          "person",
+          "person",
+          "person",
+          "person"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_ocr",
+      "name": "Extractors · OCR (LLM vision)",
+      "endpoint": "POST /v1/ocr",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 684.6,
+      "result": {
+        "text": "Asa arată\nANTI-ROMÂNISMUL\n22:08\nVOCEA LIBERTĂȚII\n15 AUG. 25\nALASKA\nTRUMP VS PUTIN\nTikTok\n@ ovi1623\n>> Da. Deci acuma nici nu mai știi cine"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_sentiment",
+      "name": "Extractors · sentiment (LLM)",
+      "endpoint": "POST /v1/sentiment",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 551.5,
+      "result": {
+        "label": "pozitiv",
+        "score": 0.9
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "forensic",
+      "name": "Forensic · module list",
+      "endpoint": "GET /api/forensic-modules",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 9.5,
+      "result": {
+        "modules": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "forensic_evidence",
+      "name": "Forensic · analiza reala",
+      "endpoint": "POST /api/forensic-evidence",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 9563.6,
+      "result": {
+        "modules_run": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ],
+        "frames": 131
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web",
+      "name": "Web · search (SearXNG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1303.6,
+      "result": {
+        "results": 3,
+        "first": "AI Deepfake Detection Online Free - Images, Videos, Voices"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web_fetch",
+      "name": "Web · fetch",
+      "endpoint": "POST /v1/fetch",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 731.4,
+      "result": {
+        "bytes": 113707
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web_gather",
+      "name": "Web · gather (evidence)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3115.4,
+      "result": {
+        "evidence_items": 5,
+        "urls_found": 5
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog",
+      "name": "Catalog · status",
+      "endpoint": "GET /catalog/v1/status",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 81.8,
+      "result": {
+        "status": "healthy",
+        "components": [
+          "llm-inference",
+          "audio-transcription",
+          "video-analysis",
+          "web-factcheck"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog_components",
+      "name": "Catalog · components",
+      "endpoint": "GET /catalog/v1/components",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 76.0,
+      "result": {
+        "components": 4
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog_models",
+      "name": "Catalog · models",
+      "endpoint": "GET /catalog/v1/models",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 67.9,
+      "result": {
+        "models": 3
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "gateway",
+      "name": "Gateway (auth+rutare)",
+      "endpoint": "GET /health + 401",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 8.8,
+      "result": {
+        "health": 200,
+        "web_no_auth": 401
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_atoms",
+      "name": "Brain · analysis_atom",
+      "endpoint": "GET /v1/analysis_atom/stats/extended",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 15.3,
+      "result": {
+        "total_atoms": 9,
+        "by_tier": {
+          "gold": 4,
+          "silver": 3,
+          "bronze": 2
+        }
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_facts",
+      "name": "Brain · fact_status",
+      "endpoint": "GET /v1/fact_status/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 14.5,
+      "result": {
+        "total": 7
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_vcache",
+      "name": "Brain · verification_cache",
+      "endpoint": "GET /v1/verification_cache/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 11.3,
+      "result": {
+        "total": 5
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_search",
+      "name": "Brain · search (RAG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3044.4,
+      "result": {
+        "total_results": 0,
+        "cache": "MISS"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_gather",
+      "name": "Brain · gather (RAG)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3038.3,
+      "result": {
+        "evidence_items": 0,
+        "urls": 0
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "dashboard",
+      "name": "Dashboard · health",
+      "endpoint": "GET /health",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 19.3,
+      "result": {
+        "status": "healthy",
+        "db": true
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "dashboard_monitoring",
+      "name": "Dashboard · monitoring 11/11",
+      "endpoint": "GET /api/monitoring/services",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 107.6,
+      "result": {
+        "healthy": "11/11"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "cloak",
+      "name": "Cloak · SERP scraper",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1991.0,
+      "result": {
+        "results": 3,
+        "first": "Deepfake - Wikipedia"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    }
+  ]
+}
--- a/ai_platform/local_gpu_stack/reports/test_report_20260627_143404.json
+++ b/ai_platform/local_gpu_stack/reports/test_report_20260627_143404.json
@ -0,0 +1,249 @@
+{
+  "generated_at": "2026-06-27T14:34:04.610734+03:00",
+  "host": "10.11.10.18",
+  "summary": {
+    "total": 17,
+    "passed": 17,
+    "failed": 0
+  },
+  "tests": [
+    {
+      "id": "llm",
+      "name": "LLM (Qwen3.5-35B-A3B)",
+      "endpoint": "POST /v1/chat/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 424.6,
+      "result": {
+        "reply": "Un deepfake este un conținut media manipulat prin tehnologia inteligenței artificiale, care înlocuiește fața sau vocea unei persoane cu cea a alteia pentru a crea imagini sau înregistrări video extrem de realiste, dar falsificate."
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "embeddings",
+      "name": "Embeddings (bge-m3)",
+      "endpoint": "POST /v1/embeddings",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 42.5,
+      "result": {
+        "dim": 1024
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "rerank",
+      "name": "Rerank (bge-reranker-v2-m3)",
+      "endpoint": "POST /v1/rerank",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 33.3,
+      "result": {
+        "top_index": 1,
+        "score": 0.626
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "audio",
+      "name": "Audio (Whisper large-v3-turbo)",
+      "endpoint": "POST /v1/audio/transcriptions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 251.9,
+      "result": {
+        "text": "Nici nu mai știi cine este de fapt agresorul și agresatul. De ce? Simplu. Pentru că în acest moment, Uniunea Europeană...",
+        "language": "ro"
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "video",
+      "name": "Video/BusterX (deepfake)",
+      "endpoint": "POST /analyze/video",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1355.9,
+      "result": {
+        "verdict": "REAL",
+        "frames_analyzed": 16
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "extractors_metadata",
+      "name": "Extractors · metadata",
+      "endpoint": "POST /v1/metadata",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 275.1,
+      "result": {
+        "media_type": "video",
+        "sha256": "27d6c00786e63905"
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "extractors_ner",
+      "name": "Extractors · NER (GLiNER)",
+      "endpoint": "POST /v1/ner",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 813.0,
+      "result": {
+        "entities": [
+          "Klaus Iohannis:persoană",
+          "Emmanuel Macron:persoană",
+          "Bucuresti:locație"
+        ]
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "extractors_detect",
+      "name": "Extractors · detect (YOLOv8)",
+      "endpoint": "POST /v1/detect",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1464.4,
+      "result": {
+        "objects": 4,
+        "labels": [
+          "person",
+          "person",
+          "person",
+          "person"
+        ]
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "extractors_ocr",
+      "name": "Extractors · OCR (LLM vision)",
+      "endpoint": "POST /v1/ocr",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 5621.1,
+      "result": {
+        "text": "Asa arată\nANTI-ROMÂNISMUL\n22:08\nVOCEA LIBERTĂȚII\n15 AUG. 25\nALASKA\nTRUMP VS PUTIN\nTikTok\n@ ovi1623\n>> Da. Deci acuma nici nu mai știi cine"
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "extractors_sentiment",
+      "name": "Extractors · sentiment (LLM)",
+      "endpoint": "POST /v1/sentiment",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 551.3,
+      "result": {
+        "label": "pozitiv",
+        "score": 0.9
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "forensic",
+      "name": "Forensic (rPPG/lip-sync/...)",
+      "endpoint": "GET /api/forensic-modules",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 6.3,
+      "result": {
+        "modules": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ]
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "web",
+      "name": "Web search (SearXNG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1504.0,
+      "result": {
+        "results": 3,
+        "first": "AI Deepfake Detection Online Free - Images, Videos, Voices"
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "catalog",
+      "name": "Catalog (via gateway)",
+      "endpoint": "GET /catalog/v1/status",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 68.9,
+      "result": {
+        "status": "healthy",
+        "components": [
+          "llm-inference",
+          "audio-transcription",
+          "video-analysis",
+          "web-factcheck"
+        ]
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "gateway",
+      "name": "Gateway (auth+rutare)",
+      "endpoint": "GET /health + 401",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 7.3,
+      "result": {
+        "health": 200,
+        "web_no_auth": 401
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "brain_atoms",
+      "name": "Brain · analysis_atom",
+      "endpoint": "GET /v1/analysis_atom/stats/extended",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 54.1,
+      "result": {
+        "total_atoms": 9,
+        "by_tier": {
+          "gold": 4,
+          "silver": 3,
+          "bronze": 2
+        }
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "brain_facts",
+      "name": "Brain · fact_status",
+      "endpoint": "GET /v1/fact_status/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 16.3,
+      "result": {
+        "total": 7
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    },
+    {
+      "id": "brain_vcache",
+      "name": "Brain · verification_cache",
+      "endpoint": "GET /v1/verification_cache/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 22.8,
+      "result": {
+        "total": 5
+      },
+      "checked_at": "2026-06-27T14:34:04.610734+03:00"
+    }
+  ]
+}
--- a/ai_platform/local_gpu_stack/reports/test_report_20260627_144358.json
+++ b/ai_platform/local_gpu_stack/reports/test_report_20260627_144358.json
@ -0,0 +1,422 @@
+{
+  "generated_at": "2026-06-27T14:43:58.666410+03:00",
+  "host": "10.11.10.18",
+  "summary": {
+    "total": 30,
+    "passed": 30,
+    "failed": 0
+  },
+  "tests": [
+    {
+      "id": "llm",
+      "name": "LLM (Qwen3.5-35B-A3B)",
+      "endpoint": "POST /v1/chat/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 400.6,
+      "result": {
+        "reply": "Deepfake este o tehnică de manipulare a conținutului multimedia care utilizează inteligența artificială pentru a înlocua fața sau vocea unei persoane reale cu cea a alteia, creând astfel imagini sau videoclipuri extrem de realiste dar false."
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "llm_completions",
+      "name": "LLM · completions",
+      "endpoint": "POST /v1/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 122.0,
+      "result": {
+        "text": " Bucuresti.\n\n<think>\n\n</think>\n\nAfirm"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "llm_info",
+      "name": "LLM · info",
+      "endpoint": "GET /v1/info",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 12.9,
+      "result": {
+        "keys": [
+          "resource",
+          "models",
+          "functions"
+        ]
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "embeddings",
+      "name": "Embeddings (bge-m3)",
+      "endpoint": "POST /v1/embeddings",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 41.8,
+      "result": {
+        "dim": 1024
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "rerank",
+      "name": "Rerank (bge-reranker-v2-m3)",
+      "endpoint": "POST /v1/rerank",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 39.7,
+      "result": {
+        "top_index": 1,
+        "score": 0.626
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "audio",
+      "name": "Audio (Whisper large-v3-turbo)",
+      "endpoint": "POST /v1/audio/transcriptions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 372.5,
+      "result": {
+        "text": "Nici nu mai știi cine este de fapt agresorul și agresatul. De ce? Simplu. Pentru că în acest moment, Uniunea Europeană...",
+        "language": "ro"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "video",
+      "name": "Video/BusterX (deepfake)",
+      "endpoint": "POST /analyze/video",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1384.5,
+      "result": {
+        "verdict": "REAL",
+        "frames_analyzed": 16
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "video_semantic",
+      "name": "Video · semantic",
+      "endpoint": "POST /analyze/video/semantic",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 27228.4,
+      "result": {
+        "num_chunks": 5,
+        "summary": "[Aggregation failed: Client error '404 Not Found' for url 'http://didiai-llm-api:14011/v1/v1/chat/completions'\nFor more "
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "extractors_metadata",
+      "name": "Extractors · metadata",
+      "endpoint": "POST /v1/metadata",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 320.3,
+      "result": {
+        "media_type": "video",
+        "sha256": "27d6c00786e63905"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "extractors_ner",
+      "name": "Extractors · NER (GLiNER)",
+      "endpoint": "POST /v1/ner",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 115.1,
+      "result": {
+        "entities": [
+          "Klaus Iohannis:persoană",
+          "Emmanuel Macron:persoană",
+          "Bucuresti:locație"
+        ]
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "extractors_detect",
+      "name": "Extractors · detect (YOLOv8)",
+      "endpoint": "POST /v1/detect",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 53.7,
+      "result": {
+        "objects": 4,
+        "labels": [
+          "person",
+          "person",
+          "person",
+          "person"
+        ]
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "extractors_ocr",
+      "name": "Extractors · OCR (LLM vision)",
+      "endpoint": "POST /v1/ocr",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 701.6,
+      "result": {
+        "text": "Asa arată\nANTI-ROMÂNISMUL\n22:08\nVOCEA LIBERTĂȚII\n15 AUG. 25\nALASKA\nTRUMP VS PUTIN\nTikTok\n@ ovi1623\n>> Da. Deci acuma nici nu mai știi cine"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "extractors_sentiment",
+      "name": "Extractors · sentiment (LLM)",
+      "endpoint": "POST /v1/sentiment",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 552.9,
+      "result": {
+        "label": "pozitiv",
+        "score": 0.9
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "forensic",
+      "name": "Forensic · module list",
+      "endpoint": "GET /api/forensic-modules",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 8.7,
+      "result": {
+        "modules": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ]
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "forensic_evidence",
+      "name": "Forensic · analiza reala",
+      "endpoint": "POST /api/forensic-evidence",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 9039.9,
+      "result": {
+        "modules_run": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ],
+        "frames": 131
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "web",
+      "name": "Web · search (SearXNG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 841.4,
+      "result": {
+        "results": 3,
+        "first": "AI Deepfake Detection Online Free - Images, Videos, Voices"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "web_fetch",
+      "name": "Web · fetch",
+      "endpoint": "POST /v1/fetch",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 554.3,
+      "result": {
+        "bytes": 113707
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "web_gather",
+      "name": "Web · gather (evidence)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 2911.9,
+      "result": {
+        "evidence_items": 5,
+        "urls_found": 5
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "catalog",
+      "name": "Catalog · status",
+      "endpoint": "GET /catalog/v1/status",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 73.8,
+      "result": {
+        "status": "healthy",
+        "components": [
+          "llm-inference",
+          "audio-transcription",
+          "video-analysis",
+          "web-factcheck"
+        ]
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "catalog_components",
+      "name": "Catalog · components",
+      "endpoint": "GET /catalog/v1/components",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 70.6,
+      "result": {
+        "components": 4
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "catalog_models",
+      "name": "Catalog · models",
+      "endpoint": "GET /catalog/v1/models",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 55.7,
+      "result": {
+        "models": 3
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "gateway",
+      "name": "Gateway (auth+rutare)",
+      "endpoint": "GET /health + 401",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 7.1,
+      "result": {
+        "health": 200,
+        "web_no_auth": 401
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "brain_atoms",
+      "name": "Brain · analysis_atom",
+      "endpoint": "GET /v1/analysis_atom/stats/extended",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 36.1,
+      "result": {
+        "total_atoms": 9,
+        "by_tier": {
+          "gold": 4,
+          "silver": 3,
+          "bronze": 2
+        }
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "brain_facts",
+      "name": "Brain · fact_status",
+      "endpoint": "GET /v1/fact_status/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 13.3,
+      "result": {
+        "total": 7
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "brain_vcache",
+      "name": "Brain · verification_cache",
+      "endpoint": "GET /v1/verification_cache/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 13.7,
+      "result": {
+        "total": 5
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "brain_search",
+      "name": "Brain · search (RAG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3040.4,
+      "result": {
+        "total_results": 0,
+        "cache": "MISS"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "brain_gather",
+      "name": "Brain · gather (RAG)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3042.0,
+      "result": {
+        "evidence_items": 0,
+        "urls": 0
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "dashboard",
+      "name": "Dashboard · health",
+      "endpoint": "GET /health",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 14.0,
+      "result": {
+        "status": "healthy",
+        "db": true
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "dashboard_monitoring",
+      "name": "Dashboard · monitoring 11/11",
+      "endpoint": "GET /api/monitoring/services",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 129.6,
+      "result": {
+        "healthy": "11/11"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    },
+    {
+      "id": "cloak",
+      "name": "Cloak · SERP scraper",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 2014.3,
+      "result": {
+        "results": 3,
+        "first": "Deepfake - Wikipedia"
+      },
+      "checked_at": "2026-06-27T14:43:58.666410+03:00"
+    }
+  ]
+}
--- a/ai_platform/local_gpu_stack/reports/test_report_20260627_144637.json
+++ b/ai_platform/local_gpu_stack/reports/test_report_20260627_144637.json
@ -0,0 +1,422 @@
+{
+  "generated_at": "2026-06-27T14:46:37.623279+03:00",
+  "host": "10.11.10.18",
+  "summary": {
+    "total": 30,
+    "passed": 30,
+    "failed": 0
+  },
+  "tests": [
+    {
+      "id": "llm",
+      "name": "LLM (Qwen3.5-35B-A3B)",
+      "endpoint": "POST /v1/chat/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 342.2,
+      "result": {
+        "reply": "Un deepfake este o media manipulat, realizat prin tehnologia inteligenței artificiale, care înlocuiează cu realism înspăimântător fața sau vocea unei persoane cu cea a altcuiva."
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "llm_completions",
+      "name": "LLM · completions",
+      "endpoint": "POST /v1/completions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 111.8,
+      "result": {
+        "text": " un oras care atrage foarte multi turisti si"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "llm_info",
+      "name": "LLM · info",
+      "endpoint": "GET /v1/info",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 14.5,
+      "result": {
+        "keys": [
+          "resource",
+          "models",
+          "functions"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "embeddings",
+      "name": "Embeddings (bge-m3)",
+      "endpoint": "POST /v1/embeddings",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 43.8,
+      "result": {
+        "dim": 1024
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "rerank",
+      "name": "Rerank (bge-reranker-v2-m3)",
+      "endpoint": "POST /v1/rerank",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 39.3,
+      "result": {
+        "top_index": 1,
+        "score": 0.626
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "audio",
+      "name": "Audio (Whisper large-v3-turbo)",
+      "endpoint": "POST /v1/audio/transcriptions",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 371.3,
+      "result": {
+        "text": "Nici nu mai știi cine este de fapt agresorul și agresatul. De ce? Simplu. Pentru că în acest moment, Uniunea Europeană...",
+        "language": "ro"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "video",
+      "name": "Video/BusterX (deepfake)",
+      "endpoint": "POST /analyze/video",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1389.8,
+      "result": {
+        "verdict": "REAL",
+        "frames_analyzed": 16
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "video_semantic",
+      "name": "Video · semantic",
+      "endpoint": "POST /analyze/video/semantic",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 18390.9,
+      "result": {
+        "num_chunks": 5,
+        "summary": "The video presents a continuous, polemical monologue delivered by a male speaker in a studio setting, likely a podcast o"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_metadata",
+      "name": "Extractors · metadata",
+      "endpoint": "POST /v1/metadata",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 284.1,
+      "result": {
+        "media_type": "video",
+        "sha256": "27d6c00786e63905"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_ner",
+      "name": "Extractors · NER (GLiNER)",
+      "endpoint": "POST /v1/ner",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 262.9,
+      "result": {
+        "entities": [
+          "Klaus Iohannis:persoană",
+          "Emmanuel Macron:persoană",
+          "Bucuresti:locație"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_detect",
+      "name": "Extractors · detect (YOLOv8)",
+      "endpoint": "POST /v1/detect",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 77.1,
+      "result": {
+        "objects": 4,
+        "labels": [
+          "person",
+          "person",
+          "person",
+          "person"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_ocr",
+      "name": "Extractors · OCR (LLM vision)",
+      "endpoint": "POST /v1/ocr",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 684.6,
+      "result": {
+        "text": "Asa arată\nANTI-ROMÂNISMUL\n22:08\nVOCEA LIBERTĂȚII\n15 AUG. 25\nALASKA\nTRUMP VS PUTIN\nTikTok\n@ ovi1623\n>> Da. Deci acuma nici nu mai știi cine"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "extractors_sentiment",
+      "name": "Extractors · sentiment (LLM)",
+      "endpoint": "POST /v1/sentiment",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 551.5,
+      "result": {
+        "label": "pozitiv",
+        "score": 0.9
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "forensic",
+      "name": "Forensic · module list",
+      "endpoint": "GET /api/forensic-modules",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 9.5,
+      "result": {
+        "modules": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "forensic_evidence",
+      "name": "Forensic · analiza reala",
+      "endpoint": "POST /api/forensic-evidence",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 9563.6,
+      "result": {
+        "modules_run": [
+          "m25",
+          "m26",
+          "m27",
+          "m28",
+          "m29"
+        ],
+        "frames": 131
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web",
+      "name": "Web · search (SearXNG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1303.6,
+      "result": {
+        "results": 3,
+        "first": "AI Deepfake Detection Online Free - Images, Videos, Voices"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web_fetch",
+      "name": "Web · fetch",
+      "endpoint": "POST /v1/fetch",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 731.4,
+      "result": {
+        "bytes": 113707
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "web_gather",
+      "name": "Web · gather (evidence)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3115.4,
+      "result": {
+        "evidence_items": 5,
+        "urls_found": 5
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog",
+      "name": "Catalog · status",
+      "endpoint": "GET /catalog/v1/status",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 81.8,
+      "result": {
+        "status": "healthy",
+        "components": [
+          "llm-inference",
+          "audio-transcription",
+          "video-analysis",
+          "web-factcheck"
+        ]
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog_components",
+      "name": "Catalog · components",
+      "endpoint": "GET /catalog/v1/components",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 76.0,
+      "result": {
+        "components": 4
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "catalog_models",
+      "name": "Catalog · models",
+      "endpoint": "GET /catalog/v1/models",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 67.9,
+      "result": {
+        "models": 3
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "gateway",
+      "name": "Gateway (auth+rutare)",
+      "endpoint": "GET /health + 401",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 8.8,
+      "result": {
+        "health": 200,
+        "web_no_auth": 401
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_atoms",
+      "name": "Brain · analysis_atom",
+      "endpoint": "GET /v1/analysis_atom/stats/extended",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 15.3,
+      "result": {
+        "total_atoms": 9,
+        "by_tier": {
+          "gold": 4,
+          "silver": 3,
+          "bronze": 2
+        }
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_facts",
+      "name": "Brain · fact_status",
+      "endpoint": "GET /v1/fact_status/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 14.5,
+      "result": {
+        "total": 7
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_vcache",
+      "name": "Brain · verification_cache",
+      "endpoint": "GET /v1/verification_cache/list",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 11.3,
+      "result": {
+        "total": 5
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_search",
+      "name": "Brain · search (RAG)",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3044.4,
+      "result": {
+        "total_results": 0,
+        "cache": "MISS"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "brain_gather",
+      "name": "Brain · gather (RAG)",
+      "endpoint": "POST /v1/gather",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 3038.3,
+      "result": {
+        "evidence_items": 0,
+        "urls": 0
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "dashboard",
+      "name": "Dashboard · health",
+      "endpoint": "GET /health",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 19.3,
+      "result": {
+        "status": "healthy",
+        "db": true
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "dashboard_monitoring",
+      "name": "Dashboard · monitoring 11/11",
+      "endpoint": "GET /api/monitoring/services",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 107.6,
+      "result": {
+        "healthy": "11/11"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    },
+    {
+      "id": "cloak",
+      "name": "Cloak · SERP scraper",
+      "endpoint": "POST /v1/search",
+      "http_code": 200,
+      "ok": true,
+      "latency_ms": 1991.0,
+      "result": {
+        "results": 3,
+        "first": "Deepfake - Wikipedia"
+      },
+      "checked_at": "2026-06-27T14:46:37.623279+03:00"
+    }
+  ]
+}
--- a/ai_platform/local_gpu_stack/run_tests.py
+++ b/ai_platform/local_gpu_stack/run_tests.py
@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""DiDi LOT 1 — runner automat de teste end-to-end.
+
+Rulează un test funcțional pe fiecare componentă și scrie rezultatele într-un
+raport JSON cu timestamp (reports/test_report_<data>.json). Fiecare intrare
+conține: componentă, endpoint, http_code, ok (verdict funcțional), latență,
+extras din răspuns + request-ul folosit — ca dovadă vizibilă.
+
+Rulare:  python3 run_tests.py            (HOST_IP din .env sau 10.11.10.18)
+         HOST_IP=1.2.3.4 python3 run_tests.py
+"""
+from __future__ import annotations
+import json, os, subprocess, sys, time, datetime, pathlib
+
+HERE = pathlib.Path(__file__).resolve().parent
+ASSETS = HERE / "test_assets"
+REPORTS = HERE / "reports"
+
+# --- config din .env (fallback la defaults) ---------------------------------
+def load_env() -> dict:
+    env = {}
+    f = HERE / ".env"
+    if f.exists():
+        for line in f.read_text().splitlines():
+            line = line.strip()
+            if line and not line.startswith("#") and "=" in line:
+                k, v = line.split("=", 1)
+                env[k.strip()] = v.strip()
+    return env
+
+ENV = load_env()
+H = os.environ.get("HOST_IP") or ENV.get("HOST_IP") or "10.11.10.18"
+TOKEN = ENV.get("GATEWAY_API_TOKEN", "didi-local-dev-token-123")
+WHISPER = ENV.get("WHISPER_MODEL", "large-v3-turbo")
+
+# --- docker shim (pt pregatirea asset-urilor; merge si prin sg docker) -------
+def docker(args: list[str]) -> subprocess.CompletedProcess:
+    if subprocess.run(["docker", "ps"], capture_output=True).returncode == 0:
+        return subprocess.run(["docker", *args], capture_output=True, text=True)
+    return subprocess.run(["sg", "docker", "-c", "docker " + " ".join(args)],
+                          capture_output=True, text=True)
+
+def ensure_assets() -> None:
+    ASSETS.mkdir(exist_ok=True)
+    mp4 = ASSETS / "sample.mp4"
+    if not mp4.exists():
+        import glob
+        cands = sorted(glob.glob(str(HERE.parent.parent / "artefacte_lot1/runs_deepfake/**/*.mp4"), recursive=True),
+                       key=lambda p: os.path.getsize(p))
+        if cands:
+            subprocess.run(["cp", cands[0], str(mp4)])
+    if not (ASSETS / "frame.jpg").exists() or not (ASSETS / "audio.wav").exists():
+        # genereaza un cadru + audio din containerul extractors (are ffmpeg)
+        docker(["cp", str(mp4), "didiAI-extractors:/tmp/s.mp4"])
+        docker(["exec", "didiAI-extractors", "sh", "-c",
+                "ffmpeg -y -i /tmp/s.mp4 -frames:v 1 /tmp/frame.jpg -vn -ar 16000 -ac 1 -t 8 /tmp/audio.wav"])
+        docker(["cp", "didiAI-extractors:/tmp/frame.jpg", str(ASSETS / "frame.jpg")])
+        docker(["cp", "didiAI-extractors:/tmp/audio.wav", str(ASSETS / "audio.wav")])
+
+# --- helper request prin curl -----------------------------------------------
+def run_curl(args: list[str], timeout: int) -> tuple[int, str, float]:
+    """Return (http_code, body, latency_ms)."""
+    full = ["curl", "-s", "-m", str(timeout), "-w", "\n%{http_code}", *args]
+    t0 = time.monotonic()
+    try:
+        p = subprocess.run(full, capture_output=True, text=True, timeout=timeout + 10)
+        out = p.stdout
+    except subprocess.TimeoutExpired:
+        return (0, "TIMEOUT", round((time.monotonic() - t0) * 1000, 1))
+    lat = round((time.monotonic() - t0) * 1000, 1)
+    code = 0; body = out
+    if "\n" in out:
+        body, _, last = out.rpartition("\n")
+        try: code = int(last.strip())
+        except ValueError: code = 0
+    return (code, body, lat)
+
+def jparse(body: str):
+    try: return json.loads(body)
+    except Exception: return None
+
+# --- definitia testelor ------------------------------------------------------
+def t_llm():
+    c, b, l = run_curl([f"http://{H}:14011/v1/chat/completions", "-H", "Content-Type: application/json",
+        "-d", json.dumps({"model":"qwen3.5","messages":[{"role":"user","content":"Ce este un deepfake? O propozitie."}],"max_tokens":120})], 90)
+    d = jparse(b); txt = (d or {}).get("choices",[{}])[0].get("message",{}).get("content","") if d else ""
+    return c, l, bool(txt) and not txt.lower().startswith("thinking"), {"reply": txt[:300]}
+def t_embeddings():
+    c,b,l = run_curl([f"http://{H}:14100/v1/embeddings","-H","Content-Type: application/json","-d",json.dumps({"model":"bge-m3","input":"test"})],30)
+    d=jparse(b); dim=len((d or {}).get("data",[{}])[0].get("embedding",[])) if d else 0
+    return c,l,dim==1024,{"dim":dim}
+def t_rerank():
+    c,b,l=run_curl([f"http://{H}:14200/v1/rerank","-H","Content-Type: application/json","-d",json.dumps({"model":"bge-reranker-v2-m3","query":"deepfake detection","documents":["o pisica","sistem detectie deepfake video"]})],30)
+    d=jparse(b); r=(d or {}).get("results",[]) ; top=r[0] if r else {}
+    return c,l,top.get("index")==1,{"top_index":top.get("index"),"score":round(top.get("relevance_score",0),4)}
+def t_audio():
+    c,b,l=run_curl([f"http://{H}:54300/v1/audio/transcriptions","-F",f"file=@{ASSETS}/audio.wav","-F",f"model={WHISPER}"],120)
+    d=jparse(b); txt=(d or {}).get("text","") if d else ""
+    return c,l,bool(txt.strip()),{"text":txt[:200],"language":(d or {}).get("language")}
+def t_video():
+    c,b,l=run_curl([f"http://{H}:54600/analyze/video","-F",f"file=@{ASSETS}/sample.mp4"],180)
+    d=jparse(b); v=(d or {}).get("verdict")
+    return c,l,v in ("REAL","FAKE","UNCERTAIN"),{"verdict":v,"frames_analyzed":(d or {}).get("frames_analyzed")}
+def t_metadata():
+    c,b,l=run_curl([f"http://{H}:54400/v1/metadata","-F",f"file=@{ASSETS}/sample.mp4"],40)
+    d=jparse(b); ok=bool((d or {}).get("analyses",{}).get("integrity",{}).get("ok"))
+    return c,l,ok,{"media_type":(d or {}).get("media_type"),"sha256":(d or {}).get("sha256","")[:16]}
+def t_ner():
+    c,b,l=run_curl([f"http://{H}:54400/v1/ner","-H","Content-Type: application/json","-d",json.dumps({"text":"Klaus Iohannis s-a intalnit cu Emmanuel Macron la Bucuresti."})],60)
+    d=jparse(b); ents=(d or {}).get("results",{}).get("entities",[])
+    return c,l,len(ents)>=2,{"entities":[e.get("text")+":"+e.get("label","") for e in ents][:5]}
+def t_detect():
+    c,b,l=run_curl([f"http://{H}:54400/v1/detect","-F",f"file=@{ASSETS}/frame.jpg"],90)
+    d=jparse(b); det=(d or {}).get("results",{}).get("detections",[])
+    return c,l,len(det)>=1,{"objects":len(det),"labels":[x.get("label") for x in det][:5]}
+def t_ocr():
+    c,b,l=run_curl([f"http://{H}:54400/v1/ocr","-F",f"file=@{ASSETS}/frame.jpg"],90)
+    d=jparse(b); ok=bool((d or {}).get("ok")); txt=(d or {}).get("results",{}).get("text","")
+    return c,l,ok,{"text":txt[:150]}
+def t_sentiment():
+    c,b,l=run_curl([f"http://{H}:54400/v1/sentiment","-H","Content-Type: application/json","-d",json.dumps({"text":"Produsul este excelent, sunt foarte multumit!"})],60)
+    d=jparse(b); ok=bool((d or {}).get("ok"))
+    return c,l,ok,{"label":(d or {}).get("results",{}).get("label"),"score":(d or {}).get("results",{}).get("score")}
+def t_forensic():
+    c,b,l=run_curl([f"http://{H}:8085/api/forensic-modules"],15)
+    d=jparse(b); m=(d or {}).get("available_modules",[])
+    return c,l,len(m)>=3,{"modules":[x.get("id") for x in m]}
+def t_web():
+    c,b,l=run_curl([f"http://{H}:51100/v1/search","-H","Content-Type: application/json","-d",json.dumps({"queries":["deepfake detection"],"max_results":3})],40)
+    d=jparse(b); r=(d or {}).get("results",[])
+    return c,l,len(r)>0,{"results":len(r),"first":(r[0].get("title","")[:60] if r else None)}
+def t_catalog():
+    c,b,l=run_curl([f"http://{H}:11000/catalog/v1/status","-H",f"Authorization: Bearer {TOKEN}"],15)
+    d=jparse(b); st=(d or {}).get("status")
+    return c,l,st in ("healthy","degraded"),{"status":st,"components":[x.get("component_id") for x in (d or {}).get("components",[])]}
+def t_gateway():
+    c,b,l=run_curl([f"http://{H}:11000/health"],10)
+    c2,_,_=run_curl([f"http://{H}:11000/web/v1/info"],10)  # fara token -> 401
+    return c,l,(c==200 and c2==401),{"health":c,"web_no_auth":c2}
+def t_brain_atoms():
+    c,b,l=run_curl([f"http://{H}:8090/v1/analysis_atom/stats/extended"],15)
+    d=jparse(b)
+    return c,l,(d is not None and "total_atoms" in d),{"total_atoms":(d or {}).get("total_atoms"),"by_tier":(d or {}).get("by_tier")}
+def t_brain_facts():
+    c,b,l=run_curl([f"http://{H}:8090/v1/fact_status/list?page=1&page_size=10"],15)
+    d=jparse(b)
+    return c,l,(d is not None and "total" in d),{"total":(d or {}).get("total")}
+def t_brain_vcache():
+    c,b,l=run_curl([f"http://{H}:8090/v1/verification_cache/list?page=1&page_size=10"],15)
+    d=jparse(b)
+    return c,l,(d is not None and "total" in d),{"total":(d or {}).get("total")}
+
+def t_llm_completions():
+    c,b,l=run_curl([f"http://{H}:14011/v1/completions","-H","Content-Type: application/json","-d",json.dumps({"model":"qwen3.5","prompt":"Capitala Romaniei este","max_tokens":10})],60)
+    d=jparse(b); txt=(d or {}).get("choices",[{}])[0].get("text","") if d else ""
+    return c,l,bool(txt.strip()),{"text":txt[:120]}
+def t_llm_info():
+    c,b,l=run_curl([f"http://{H}:14011/v1/info"],15); d=jparse(b)
+    return c,l,isinstance(d,dict) and len(d)>0,{"keys":list(d.keys())[:6] if isinstance(d,dict) else None}
+def t_video_semantic():
+    c,b,l=run_curl([f"http://{H}:54600/analyze/video/semantic","-F",f"file=@{ASSETS}/sample.mp4"],180)
+    d=jparse(b); ok=bool(d) and ("final_summary" in d or "chunk_results" in d)
+    return c,l,ok,{"num_chunks":(d or {}).get("num_chunks"),"summary":(str((d or {}).get("final_summary",""))[:120])}
+def t_web_fetch():
+    c,b,l=run_curl([f"http://{H}:51100/v1/fetch","-H","Content-Type: application/json","-d",json.dumps({"urls":["https://en.wikipedia.org/wiki/Deepfake"]})],60)
+    ok=(c==200 and len(b)>1500 and ("title" in b or "content" in b or "text" in b))
+    return c,l,ok,{"bytes":len(b)}
+def t_web_gather():
+    c,b,l=run_curl([f"http://{H}:51100/v1/gather","-H","Content-Type: application/json","-d",json.dumps({"claim":"Deepfakes can be detected by AI","max_search_results":5})],150)
+    d=jparse(b); ev=(d or {}).get("evidence",[]); n=(d or {}).get("total_evidence_items",len(ev) if isinstance(ev,list) else 0)
+    return c,l,bool(d) and (n>0 or (isinstance(ev,list) and len(ev)>0)),{"evidence_items":n,"urls_found":(d or {}).get("total_urls_found")}
+def t_forensic_evidence():
+    c,b,l=run_curl([f"http://{H}:8085/api/forensic-evidence","-F",f"video=@{ASSETS}/sample.mp4","-F","modules=m29","-F","every_n_frames=30"],150)
+    d=jparse(b); mr=(d or {}).get("modules_run",[])
+    return c,l,bool(mr),{"modules_run":mr,"frames":(d or {}).get("n_frames_extracted")}
+def t_catalog_components():
+    c,b,l=run_curl([f"http://{H}:11000/catalog/v1/components","-H",f"Authorization: Bearer {TOKEN}"],15)
+    d=jparse(b); comp=(d or {}).get("components",d if isinstance(d,list) else [])
+    return c,l,c==200 and len(comp)>0,{"components":len(comp)}
+def t_catalog_models():
+    c,b,l=run_curl([f"http://{H}:11000/catalog/v1/models","-H",f"Authorization: Bearer {TOKEN}"],15)
+    d=jparse(b)
+    return c,l,c==200 and d is not None,{"models":len((d or {}).get("models",d if isinstance(d,list) else []))}
+def t_brain_search():
+    c,b,l=run_curl([f"http://{H}:8090/v1/search","-H","Content-Type: application/json","-d",json.dumps({"queries":["deepfake"],"max_results":2})],60)
+    d=jparse(b)
+    return c,l,bool(d) and "results" in d,{"total_results":(d or {}).get("total_results"),"cache":(d or {}).get("brain_meta",{}).get("cache_status")}
+def t_brain_gather():
+    c,b,l=run_curl([f"http://{H}:8090/v1/gather","-H","Content-Type: application/json","-d",json.dumps({"claim":"Deepfakes can be detected by AI"})],120)
+    d=jparse(b)
+    return c,l,bool(d) and "evidence" in d,{"evidence_items":(d or {}).get("total_evidence_items"),"urls":(d or {}).get("total_urls_found")}
+def t_dashboard_health():
+    c,b,l=run_curl([f"http://{H}:51300/health"],10); d=jparse(b)
+    return c,l,(d or {}).get("status") in ("healthy","ok"),{"status":(d or {}).get("status"),"db":(d or {}).get("db")}
+def t_dashboard_monitoring():
+    c,b,l=run_curl([f"http://{H}:51300/api/monitoring/services"],20); d=jparse(b)
+    s=d if isinstance(d,list) else (d or {}).get("services",[])
+    h=sum(1 for x in s if x.get("status")=="healthy")
+    return c,l,len(s)>0 and h==len(s),{"healthy":f"{h}/{len(s)}"}
+def t_cloak():
+    c,b,l=run_curl([f"http://{H}:8770/v1/search","-H","Content-Type: application/json","-d",json.dumps({"queries":["deepfake"],"engines":["ddg"],"max_results_per_engine":3})],45)
+    d=jparse(b); r=(d or {}).get("results",[])
+    return c,l,len(r)>0,{"results":len(r),"first":(r[0].get("title","")[:55] if r else None)}
+
+TESTS = [
+    ("llm","LLM (Qwen3.5-35B-A3B)","POST /v1/chat/completions",t_llm),
+    ("llm_completions","LLM · completions","POST /v1/completions",t_llm_completions),
+    ("llm_info","LLM · info","GET /v1/info",t_llm_info),
+    ("embeddings","Embeddings (bge-m3)","POST /v1/embeddings",t_embeddings),
+    ("rerank","Rerank (bge-reranker-v2-m3)","POST /v1/rerank",t_rerank),
+    ("audio","Audio (Whisper large-v3-turbo)","POST /v1/audio/transcriptions",t_audio),
+    ("video","Video/BusterX (deepfake)","POST /analyze/video",t_video),
+    ("video_semantic","Video · semantic","POST /analyze/video/semantic",t_video_semantic),
+    ("extractors_metadata","Extractors · metadata","POST /v1/metadata",t_metadata),
+    ("extractors_ner","Extractors · NER (GLiNER)","POST /v1/ner",t_ner),
+    ("extractors_detect","Extractors · detect (YOLOv8)","POST /v1/detect",t_detect),
+    ("extractors_ocr","Extractors · OCR (LLM vision)","POST /v1/ocr",t_ocr),
+    ("extractors_sentiment","Extractors · sentiment (LLM)","POST /v1/sentiment",t_sentiment),
+    ("forensic","Forensic · module list","GET /api/forensic-modules",t_forensic),
+    ("forensic_evidence","Forensic · analiza reala","POST /api/forensic-evidence",t_forensic_evidence),
+    ("web","Web · search (SearXNG)","POST /v1/search",t_web),
+    ("web_fetch","Web · fetch","POST /v1/fetch",t_web_fetch),
+    ("web_gather","Web · gather (evidence)","POST /v1/gather",t_web_gather),
+    ("catalog","Catalog · status","GET /catalog/v1/status",t_catalog),
+    ("catalog_components","Catalog · components","GET /catalog/v1/components",t_catalog_components),
+    ("catalog_models","Catalog · models","GET /catalog/v1/models",t_catalog_models),
+    ("gateway","Gateway (auth+rutare)","GET /health + 401",t_gateway),
+    ("brain_atoms","Brain · analysis_atom","GET /v1/analysis_atom/stats/extended",t_brain_atoms),
+    ("brain_facts","Brain · fact_status","GET /v1/fact_status/list",t_brain_facts),
+    ("brain_vcache","Brain · verification_cache","GET /v1/verification_cache/list",t_brain_vcache),
+    ("brain_search","Brain · search (RAG)","POST /v1/search",t_brain_search),
+    ("brain_gather","Brain · gather (RAG)","POST /v1/gather",t_brain_gather),
+    ("dashboard","Dashboard · health","GET /health",t_dashboard_health),
+    ("dashboard_monitoring","Dashboard · monitoring 11/11","GET /api/monitoring/services",t_dashboard_monitoring),
+    ("cloak","Cloak · SERP scraper","POST /v1/search",t_cloak),
+]
+
+def main() -> int:
+    ensure_assets()
+    ts = datetime.datetime.now().astimezone()
+    stamp = ts.strftime("%Y%m%d_%H%M%S")
+    results = []
+    print(f"\n  DiDi LOT 1 — test run @ {ts.isoformat()}  (host {H})\n")
+    for tid, name, ep, fn in TESTS:
+        try:
+            code, lat, ok, extra = fn()
+        except Exception as e:  # noqa: BLE001
+            code, lat, ok, extra = 0, 0.0, False, {"error": str(e)}
+        mark = "\033[32mPASS\033[0m" if ok else "\033[31mFAIL\033[0m"
+        print(f"  [{mark}] {name:34} {ep:38} HTTP {code:>3}  {lat:>7.0f}ms  {json.dumps(extra, ensure_ascii=False)[:80]}")
+        results.append({"id":tid,"name":name,"endpoint":ep,"http_code":code,
+                        "ok":ok,"latency_ms":lat,"result":extra,"checked_at":ts.isoformat()})
+    passed = sum(1 for r in results if r["ok"])
+    report = {
+        "generated_at": ts.isoformat(),
+        "host": H,
+        "summary": {"total": len(results), "passed": passed, "failed": len(results) - passed},
+        "tests": results,
+    }
+    REPORTS.mkdir(exist_ok=True)
+    out = REPORTS / f"test_report_{stamp}.json"
+    out.write_text(json.dumps(report, indent=2, ensure_ascii=False))
+    latest = REPORTS / "latest.json"
+    latest.write_text(json.dumps(report, indent=2, ensure_ascii=False))
+    print(f"\n  === {passed}/{len(results)} PASS ===")
+    print(f"  Raport JSON: {out}")
+    print(f"  (si copie: {latest})\n")
+    return 0 if passed == len(results) else 1
+
+if __name__ == "__main__":
+    sys.exit(main())
--- a/ai_platform/local_gpu_stack/seed/seed_brain.sh
+++ b/ai_platform/local_gpu_stack/seed/seed_brain.sh
@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Brain stack helpers: render env, align postgres password, create atomic token,
+# seed the canonical taxonomy. Idempotent. Requires lib.sh sourced + master .env
+# loaded ($POSTGRES_PASSWORD, $ATOMIC_TOKEN, $HOST_IP, $ENV_FILE, $BRAIN_DIR).
+
+# Write didi_brain/.env (read by brain-api/scheduler) + infra/.env (compose interpolation)
+brain_write_env() {
+  local pw=$1 tok=$2
+  cp "${BRAIN_DIR}/.env.example" "${BRAIN_DIR}/.env"
+  sed -i \
+    -e "s#^LLM_ROUTER_URL=.*#LLM_ROUTER_URL=http://didiAI-llm-api:14011#" \
+    -e "s#^LLM_VLLM_URL=.*#LLM_VLLM_URL=http://didiAI-vllm-qwen3.5:14001#" \
+    -e "s#^EMBEDDING_URL=.*#EMBEDDING_URL=http://didiAI-embeddings-api:14100#" \
+    -e "s#^RERANKER_URL=.*#RERANKER_URL=http://didiAI-rerank-api:14200#" \
+    -e "s#^MODEL_FAST_ENABLED=.*#MODEL_FAST_ENABLED=true#" \
+    -e "s#^POSTGRES_PASSWORD=.*#POSTGRES_PASSWORD=${pw}#" \
+    "${BRAIN_DIR}/.env"
+  if grep -q '^ATOMIC_TOKEN=' "${BRAIN_DIR}/.env"; then
+    sed -i "s#^ATOMIC_TOKEN=.*#ATOMIC_TOKEN=${tok}#" "${BRAIN_DIR}/.env"
+  else
+    echo "ATOMIC_TOKEN=${tok}" >> "${BRAIN_DIR}/.env"
+  fi
+  printf 'POSTGRES_USER=atomic\nPOSTGRES_PASSWORD=%s\nPOSTGRES_DB=atomic\n' "$pw" > "${BRAIN_DIR}/infra/.env"
+}
+
+# Force the postgres role password to match (volume may have been initialized with a different one)
+brain_align_pg() {
+  local pw=$1
+  docker exec didibrain-postgres psql -U atomic -d atomic \
+    -c "ALTER USER atomic WITH PASSWORD '${pw}';" >/dev/null 2>&1 || true
+}
+
+# Create an atomic API token if the master .env doesn't have one yet. Echos the token.
+brain_ensure_token() {
+  local pw=$1 cur=$2
+  if [ -n "$cur" ]; then printf '%s' "$cur"; return 0; fi
+  local out tok
+  out=$(docker exec didibrain-atomic atomic-server token create --name didibrain \
+          --storage postgres --database-url "postgres://atomic:${pw}@postgres:5432/atomic" 2>&1)
+  tok=$(printf '%s' "$out" | grep -oE 'at_[A-Za-z0-9_.-]{20,}' | head -1)
+  printf '%s' "$tok"
+}
+
+# Seed canonical taxonomy into atomic (idempotent; tags created if missing).
+brain_seed_taxonomy() {
+  docker cp "${BRAIN_DIR}/scripts" didibrain-api:/app/scripts >/dev/null 2>&1 || true
+  # local _tag_ids.json write is read-only in the image — harmless, atomic gets the tags
+  docker exec -e PYTHONPATH=/app didibrain-api python /app/scripts/04_seed_taxonomy.py >/dev/null 2>&1 || true
+}
--- a/ai_platform/local_gpu_stack/seed/seed_catalog.sh
+++ b/ai_platform/local_gpu_stack/seed/seed_catalog.sh
@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+# Idempotent seed of the dashboard DB-backed Model & Extractor catalog.
+# Defines seed_catalog(); requires lib.sh sourced + $HOST_IP set.
+
+seed_catalog() {
+  local base="http://${HOST_IP}:51300/api/catalog"
+  # existing names (idempotency)
+  local existing
+  existing=$(curl -s -m10 "$base" | python3 -c 'import sys,json
+try:
+ d=json.load(sys.stdin); items=d if isinstance(d,list) else d.get("items",d.get("data",[]))
+ print("\n".join(x.get("name","") for x in items))
+except Exception: pass' 2>/dev/null)
+
+  local entries=(
+'{"kind":"model","name":"qwen3.5","display_name":"Qwen3.5-35B-A3B","service":"llm","capabilities":["chat","completion","reasoning"],"context_length":32000,"supports_gpu":true,"endpoint":"http://HOST:14011","notes":"vLLM MoE, GPU 0"}'
+'{"kind":"model","name":"bge-m3","display_name":"BGE-M3 Embeddings","service":"embeddings","capabilities":["embeddings"],"context_length":8192,"supports_gpu":true,"endpoint":"http://HOST:14100","notes":"1024-dim, vLLM GPU 1"}'
+'{"kind":"model","name":"bge-reranker-v2-m3","display_name":"BGE Reranker v2-m3","service":"rerank","capabilities":["rerank"],"context_length":8192,"supports_gpu":true,"endpoint":"http://HOST:14200","notes":"cross-encoder, vLLM GPU 1"}'
+'{"kind":"model","name":"busterx","display_name":"BusterX++ (deepfake)","service":"video","capabilities":["deepfake-detection","video-vlm"],"context_length":32768,"supports_gpu":true,"endpoint":"http://HOST:54600","notes":"Qwen2.5-VL based, vLLM GPU 1"}'
+'{"kind":"model","name":"whisper-large-v3-turbo","display_name":"Whisper large-v3-turbo","service":"audio","capabilities":["transcription","stt"],"supports_gpu":true,"endpoint":"http://HOST:54300","notes":"faster-whisper, GPU 0"}'
+'{"kind":"extractor","name":"gliner-multi-v2.1","display_name":"GLiNER NER (multilingual)","service":"extractors","capabilities":["ner"],"supports_gpu":true,"endpoint":"http://HOST:54400/v1/ner","notes":"validat RO"}'
+'{"kind":"extractor","name":"yolov8n","display_name":"YOLOv8 Object Detection","service":"extractors","capabilities":["object-detection"],"supports_gpu":true,"endpoint":"http://HOST:54400/v1/detect"}'
+'{"kind":"extractor","name":"metadata","display_name":"Media Metadata (EXIF/ELA/ffprobe/C2PA)","service":"extractors","capabilities":["exif","ela","spectrogram","integrity"],"supports_cpu":true,"endpoint":"http://HOST:54400/v1/metadata"}'
+'{"kind":"extractor","name":"forensic-suite","display_name":"Forensic detectors (rPPG/lip-sync/forgery/lighting)","service":"forensic","capabilities":["rppg","lip-sync","forgery-heatmap","lighting"],"supports_gpu":true,"endpoint":"http://HOST:8085/api/forensic-evidence"}'
+  )
+  local added=0 skipped=0 e name
+  for e in "${entries[@]}"; do
+    name=$(printf '%s' "$e" | python3 -c 'import sys,json;print(json.load(sys.stdin)["name"])')
+    if printf '%s\n' "$existing" | grep -qx "$name"; then skipped=$((skipped+1)); continue; fi
+    e=${e//HOST/$HOST_IP}
+    if [ "$(curl -s -m10 -o /dev/null -w '%{http_code}' -X POST "$base" -H 'Content-Type: application/json' -d "$e")" = "200" ]; then
+      added=$((added+1))
+    fi
+  done
+  echo "catalog: +$added adaugate, $skipped existau deja"
+}