MD SAH ALAM BASITH

PathLens AI — Evidence-Aware Accessibility Intelligence. Solo-built for the Meta Global AI Developer Hackathon 2026, Indie track. Live: https://pathlens.cognitivebd.com · Video: https://youtu.be/6IHM7WiYKSg Problem: a wheelchair user cannot tell from a normal map whether a footpath is usable right now — is a ramp blocked, is an old obstacle still there, how old is the info, is another route better supported, or is there simply no information at all. What it does: anyone reports an obstacle with GPS/map pin (no login), uploads a photo, and runs Meta AI. `muse-spark-1.3` classifies the photo (type + confidence + visible evidence + limitations, never a verdict) and `sam-3.1` measures the occupied share of the visible walkway. The same place's older and newer photos are compared into 4 explained outcomes — LikelyStillPresent, NotObservedInNewImage (never means removed), ConflictingEvidence, Inconclusive. Reviewers verify/dispute/reject with a note; every decision is stored beside the AI outcome with actor, previous value and time. Routes use real OSRM foot geometry ranked by fewest known obstacles, with unknown share, freshness, verification split and stair warnings — unknown stays unknown, and OSRM foot is disclaimed as not wheelchair-validated, so no route is ever labelled safe. Live traffic-camera feeds (SF only) poll into the same pipeline: unchanged frame creates nothing, changed frame becomes evidence or an auto CameraDetected candidate. Accuracy (live, all Meta): SF v2 6/7 + World20 12/12 = 18/19 answerable (94.7%), 0 errors, 0 upload failures. V1 3/3/2/0 on 10 kept as the run record beside the fix (labels corrected from photos, temperature 0, abstain guardrail 0.35, fair Unknown rule). Detail: docs/AI_EVALUATION_SF_REAL_V2.md + docs/WORLD20_BATCH12.md. Built solo: Next.js + ASP.NET Core 10 + PostGIS + OSM/Leaflet + JWT review auth + CI/CD + /manual + ET timestamps. Unknown is unknown.