Machine learning · Crisis response · 2026
ReliefLens AI
An offline-first workspace that turns multilingual crisis reports into prioritized, explainable action cards.
Role · Product design, ML & full-stack development
Stack · TypeScript, Vite, Naive Bayes, PWA, Vitest
01 · The challenge
Small disaster-response teams receive urgent reports across SMS, hotline, and chat in multiple languages, often with repeated or incomplete information that slows down prioritization.
02 · The response
I built an offline-first, human-in-the-loop triage workspace with a browser-trained Naive Bayes classifier, urgency scoring, entity extraction, and character-trigram duplicate detection. All inference runs locally so sensitive messages stay on the device.
03 · What I built
English, Hindi & Bengali analysis
Explainable priority and duplicate scoring
Tested, installable offline PWA
Next case study