Trail Sentinel

Trail Sentinel is a real-time trail hazard HUD that runs in any phone or desktop browser. Point the camera down a trail and it finds hazards, ranks them, and speaks a short directional alert ("Snake, 11 o'clock, immediate") only when it matters. HONEST BACKEND STATUS (please read first) - Meta SAM 3.1 integration: implemented (meta_transport.py, sam_client.py) and verified with one live Meta Model API segment-image call on the Meta hackathon VM on 2026-10-31 (prompt "snake" on a copperhead still; one box plus pixel mask, ~3.0 s, USD 0.0025). The verbatim recorded response is in tests/fixtures/sam_hero_snake_raw.txt and decodes with Meta's public meta-sam-parser; the video shows that real mask, labelled as a recording. - Public Cloud Run demo: it is keyless (no Meta credential in the container), so it cannot call the Meta Model API. It runs a Google Vertex AI (Gemini) vision fallback and draws a bounding-box-derived contour tint, not a SAM mask. The model names in the HUD switcher (SAM 3.1 + Muse Spark, Wonder, Glimmer) describe the intended Meta route; the HUD now has an "Executed Backend" row and notice showing which backend actually answered. The README repeats this. THE PROBLEM Hikers, trail runners and cyclists get bitten, charged or injured by hazards they never saw: a camouflaged snake in leaf litter, a boar around the bend, a bear beside the path. Generic detectors cover a fixed label list and cannot be told about a new hazard on the spot. WHAT IS BUILT AND WORKS TODAY - Triage engine (deterministic, tested): severity x proximity x approach speed, alert radii 3.0 / 2.5 / 1.0 m, 3 s anti-chatter cooldown, so the hiker hears one useful alert, not noise. - Honesty by design: results older than 3.0 s are discarded and flagged "unverified"; stub/demo/backend badges; assistive-tool disclaimer. It is an assistive tool, not a safety guarantee. - Zero-retraining concepts: type or speak a new hazard ("fallen tree") and POST /api/concepts adds it to the next scan request (a SAM text prompt on the Meta route). - Localized alerts in English, Korean and Japanese: UI plus 126 pre-generated neural voice clips, haptic vibration on mobile, QR for instant phone access. - Engineering: zero-dependency Python server with a daily CallBudget cap, 78 unit tests, an evaluation harness (detection rate, lead time, false alarms), a 10-check verify_all gate, keyless Cloud Run deployment. MIT licensed. TRY IT Live demo (Vertex fallback, tap "Scan Verification Video"): https://trail-sentinel-841059355288.asia-northeast1.run.app Code: https://github.com/buddypia/trail-sentinel Built solo during the hackathon. Test footage is rights-cleared (Pexels License / Wikimedia Commons CC); attributions in eval/data/SOURCES.md.