Epic: On-device acute fall/crash detector — synthetic data -> GPU training -> tiny on-device model

[nandal]

What

The fast-acute path of the detection brain: a tiny on-device model that fuses phone sensors and decides whether a hard event — a bike crash, a fall on the stairs, a collapse — just happened, opening the "are you okay?" grace (~10–30s) before summoning help. Inference runs on-device (the raw trace never leaves the edge); training happens off-device on GPUs.

Part of the detection engine (#28, #37); fast-acute depth relates to the wearable (#8). Design discussion + trade-offs: #58.

The loop

OFF-DEVICE (GPUs, one-time)                 ON-DEVICE (always-on)
generate synthetic data ─┐
public datasets ─────────┤→ train/tune → quantise/convert → push ─► load once,
                         │   tiny model   (TFLite/CoreML/      (bundled    run inference:
                         │                 ONNX/ExecuTorch)     + OTA)     sensors in → verdict out
                  benchmark gate ◄─────────────────────────────┘

The phone only ever runs inference; it never trains.

Foundation — open now (independent, parallelisable)

• feat(sim): synthetic IMU sensor-data generator for fall/crash scenarios (domain-randomised) #59 — Synthetic IMU sensor-data generator (domain-randomised)
• feat(sim): fall/crash detection benchmark — fixed scenario suite + recall/false-alarm metrics #60 — Detection benchmark + recall/false-alarm metrics
• feat(data): adapters for public fall/HAR datasets (SisFall, UMAFall, MobiAct) -> common format #61 — Public fall/HAR dataset adapters (SisFall / UMAFall / MobiAct)

Next — promote to issues once the foundation lands (or someone's circling)

• Baseline tiny model + GPU training script + model card
• On-device inference behind a LivenessSource — emit the acute verdict, never a trace
• Model packaging: versioned artifact, bundled baseline (Sovereign-safe, no backend) + benchmark-gated OTA update with rollback

Guardrails (non-negotiable — see CLAUDE.md)

• Edge-only: on-device inference, one-bit verdict out, no telemetry logged or synced.
• Honest metrics: every model ships with a model card stating its train/test distribution; an OTA push must beat the current model on recall AND false-alarm rate, staged and rollback-able. A silently-pushed worse model is a safety regression.
• The grace window does the safety work — the detector opens a check; it does not summon help by itself. False alarm > no alarm, but keep it cheap.

Where to start

New contributors: pick up #59, #60, or #61 — they need no GPU or ML depth and unlock everything downstream. Debate architecture and datasets in #58.

[wrokin]

🤖 wrokin triage

Epic to develop an on-device acute fall/crash detector using synthetic data, GPU training, and a tiny on-device model with strict privacy and safety guardrails.

Suggested priority: high
Labels applied: enhancement
Possible duplicate: no

_{Model: deepseek/deepseek-v4-pro · your key, your model (BYOK)}

[bytesbrains]

Design note — fold GPS location in as a confidence signal (edge-only)

Beyond speed (already noted in #59), where an event happens is itself evidence. Known accident-prone locations — junctions, sharp bends, documented crash black-spots — raise the prior for a real acute event: a hard deceleration + tumble at a known black-spot is far more likely a genuine crash than the same signature in a living room. So proximity to a hotspot is a useful context feature that both boosts confidence on real events and helps trim false alarms elsewhere.

Guardrail — this must stay edge-only (non-negotiable, CLAUDE.md):

• The hotspot map is shipped to the device (bundled / cached) and looked up on-device.
• Only a derived feature feeds the model — e.g. "near a known high-risk location: yes/no" or a risk weight. The raw GPS trace is never logged or sent off the phone.
• Liveness is one bit, processed at the edge; iEye watches over you, never watches you. We ship the map out, never the user's location.

Plumbing impact: the synthetic generator (#59) and benchmark (#60) should carry a location_risk context feature so models can learn to use it; sources worth a look are public road-accident / black-spot datasets (national & city open-data crash maps; India: MoRTH / iRAD black-spot data).