ROADMAP.md

Sentinel — Roadmap

Started from a broken single-camera MVP script. Building it up in phases, each one shipping something that runs.

Phase	Layer	What	Status
0	—	Original MVP	Broken — fixed in 2.0
2.0	—	Runnable single-camera pipeline (source → detect → track → record → sign → backend)	Done
2.1	L1	Perception core — trained detector (YOLO12), RF-DETR bake-off path, training scripts	Mostly done — detector trained and live; package class and quantization still open
2.2	L5	Evidence hardening — Merkle anchor, chain of custody, OpenTimestamps	Done
2.3	L4	Edge resilience — store-and-forward outbox, heartbeat/silent-site detection	Done
2.4	L2	Reasoning — context rules, event descriptions (template + local VLM)	Done for the free tier; frontier VLM is a stub pending an API key
2.5	L6	Platform — live dashboard, SSE push, notifications, multi-tenant	Done
2.6	L3	Learning — federated learning simulation	Simulation done; real fleet-scale learning needs an actual fleet

L1 perception fleet (pose/fall, re-ID, anomaly, audio, fire/smoke) — all shipped with free, tested defaults; heavier models are a documented upgrade path once there's a reason to swap.

See PROJECT_STATUS.md for the detailed breakdown of what's tested vs. what's blocked on real hardware/data/an account.

Next up (needs you, not more code)

1. Run the Colab bake-off (RF-DETR vs YOLO12) for real
2. Pick a VLM tier — local Qwen setup or a frontier API key
3. Get package class training data
4. Pick edge hardware
5. Deploy to a real site