gridClear — models service

CCTV crash-scene oracle before arrival. A single still (or a few ESP32-CAM frames) goes in; structured findings, a pre-arrival briefing, response decisions, and a bounding-box overlay come out. Runs on the DGX Spark against the local Ollama.

Pipeline

ingest frame(s) ─► Nemotron-VL vision pass ─► evidence gate ─► triage ─► decisions + briefing
                         (structured JSON,        (damage/debris/   (confidence +
                          confidence + bbox)        ped required)      confirm banner)
                                                                              │
                              bbox overlay ◄───────────────────────────────┘
                              optional multi-image → 3D (VGGT)
                              raw photos removed (ESP32 signal contract)

Models (on the DGX Spark, via Ollama)

Task	Model	Backend
Vision (structured findings + bbox)	nemotron3:33b (Nemotron, multimodal)	Ollama /api/chat, JSON-schema constrained
Triage / confidence gate	nemotron3:33b (configurable → a dedicated Nemotron Nano)	Ollama /api/chat
Multi-image → 3D (optional)	VGGT (feed-forward, unposed)	PyTorch cu130

API

• POST /v1/scene/ingest — upload frame(s) → {session_id, n_images}
• POST /v1/scene/{session_id}/signal — ESP32 "done" signal → runs pipeline, removes photos → SceneReport
• POST /v1/scene/analyze — ingest + signal in one call (demo/tests) → SceneReport
• GET /v1/scene/{session_id}/overlay — annotated still (JPEG)
• GET /healthz, GET /v1/models/status

Run

python3 -m venv .venv && ./.venv/bin/pip install -r requirements.txt
cp .env.example .env          # adjust if needed
./scripts/run.sh              # serves on 0.0.0.0:8088

Test

./.venv/bin/python -m pytest          # fast, mocked (no GPU)
GRIDCLEAR_LIVE=1 ./.venv/bin/python -m pytest tests/test_live_dgx.py  # real models on the DGX
# optional: GRIDCLEAR_TEST_IMAGE=/path/to/crash.jpg for a real still

Config is env-driven (prefix GRIDCLEAR_); see .env.example. Set GRIDCLEAR_ENABLE_3D=true once torch+vggt are installed to turn on reconstruction.