QWENCLOUD HACKATHON · TRACK 5 — EDGEAGENT
A $60 edge guardian for clinic cold chains. Local reflexes fire in sub-milliseconds; a Qwen cloud brain diagnoses each excursion like a refrigeration engineer — "defrost cycle, benign" vs "door ajar, stock at risk in 22 minutes" — and a SHA-256 hash-chained, Ed25519-signed log keeps compliance tamper-evident through blackouts.
The magic moment
Real output from permafrost replay — the network is cut mid-run at tick 1700. The local reflex layer keeps protecting the fridge, events queue ECIES-sealed, and on reconnect the hash chain closes gap-free.
The cloud disappears mid-excursion. No panic, no gap — local rules v1 take over.
The door-ajar reflex trips locally in <1 ms, with zero connectivity.
The ECIES-sealed queue drains; verdicts arrive from the Qwen brain.
verify-chain re-derives every hash — the blackout leaves no hole in the audit log.
Reproduce it in CI form: python scripts/verify_offline.py kills real sockets mid-replay and exits 0.
What ships
A 10-second sampler feeds a local rules engine that trips the buzzer at 0.007 ms p95 — four orders of magnitude inside the 100 ms budget, and always before any cloud round-trip. The fridge is protected even if the internet never existed.
Two engineered curves spike +3.2 °C almost identically. Qwen reads the co-signals and calls one benign (defrost) and one critical (door ajar) — killing the false-alarm fatigue that gets monitors unplugged.
The cloud distills its own verdict history into local rules, signs the bundle, and the edge verifies the signature before hot-swapping. Unsigned, forged, or downgraded bundles are refused — and the refusal is chain-logged.
Excursion events — never raw streams — are ECIES-sealed to the cloud's public key and queued through outages. On reconnect the queue drains and verdicts land, with the chain gap-free across the blackout.
Every entry extends hᵢ = SHA256(hᵢ₋₁ ‖ entry) with signed daily Merkle roots. Flip one byte anywhere and verify-chain fails, names the sequence, and exits 1 — tested across all four curves.
Honest by design
These lines are quoted verbatim from the repo's README — the project states its own limits in-product.
The live model path was smoke-tested with a real DashScope call, so the wiring is verified — what this build does not contain is a full end-to-end captured live run; the deployed endpoints and judging path run the offline FakeQwen engine by design.README.md · Honesty note — a bogus key still proves the DashScope wiring with a real 401 request_id, keylessly
Real GPIO/hardware test — no physical rig has run yet. Replay mode is the supported judging path.README.md · Status (honest) — the Pi is showmanship; the loop is the product
Physical sensor spoofing (a heat gun on a probe) is out of scope; curve-only inference can flag a failing compressor but cannot confirm it without refrigerant telemetry — verdicts say so.README.md · Residual risk — in-product honesty
Questions
PERMAFROST
TRACK 5 · EDGEAGENT
FakeQwen, a deterministic fixture-backed transport. Live Qwen is an explicit opt-in: set DASHSCOPE_API_KEY and PERMAFROST_LIVE=1 and the identical command routes through real DashScope.scripts/verify_offline.py proves it in CI by killing real sockets mid-replay.qwen3.7-plus, qwen3.6-flash, text-embedding-v4) is wired end-to-end — a bogus key drives a real DashScope round-trip and returns an authentic 401 with a real request_id, proving the endpoint is no stub.verify-chain fails and exits 1 — across all four seed curves, and gap-free through a simulated power cut (os._exit mid-replay).pip install -e ".[dev]" then permafrost replay --curve seeds/door_ajar.csv --db audit.db --fresh. Follow with permafrost verify-chain audit.db and pytest for the full 331-test suite. The complete judge script is in DEMO.md.Clone it, replay the door-ajar curve, watch the reflex fire before the cloud answers — then try to tamper with the log.