Hackathon

Tanbih

A hackathon-winning crisis alert system that became an internship project, focused on the backend evidence and review workflow behind public notifications.

Timeline: March to April 2026

Technologies used

Node.js
Express
Socket.IO
Sequelize
SQLite
Gemini
Perplexity
Web Push

The backend reports source availability so an analyst can see which providers will run before detection.

Tanbih control panel showing selectable crisis data sources and unavailable credentialed providers — The backend reports source availability so an analyst can see which providers will run before detection.

A human review gate separates detected findings from multilingual public delivery.

Tanbih analyst dashboard reviewing a synthetic earthquake alert with verification confidence and recipient zones — A human review gate separates detected findings from multilingual public delivery.

The stored verification result explains how each evidence class contributed to the final score.

Tanbih verification breakdown showing semantic, scientific, and media contributions to a synthetic alert score — The stored verification result explains how each evidence class contributed to the final score.

From a hackathon win to an internship project

Tanbih started as the project that won a hackathon and earned me an internship. We continued developing it during that internship, moving beyond the original demonstration toward a more deliberate backend for crisis detection, verification, and alert review.

My work concentrated on the evidence pipeline: choosing which providers should run, keeping their outputs traceable, verifying a finding through several evidence classes, and presenting the result to a human analyst before any public notification.

Source selection is part of the record

The Express backend can ingest scientific feeds, weather services, news providers, social sources, and Lebanese public information feeds. Each run records requested, executed, unavailable, and ignored sources. The control panel reflects backend availability, so a missing credential is visible before a detection run rather than silently reducing coverage.

Verification that can be inspected

A finding is enriched with scientific confirmation, news and web corroboration, semantic checks through Gemini or Perplexity, and optional media evidence. Uploaded images can contribute metadata, visual analysis, and reverse-search results. The final record stores the contribution of each evidence class instead of exposing only a single confidence number.

A person remains between confidence and delivery

Verified findings enter an analyst queue. The dashboard shows severity, radius, confidence, evidence, recipient zones, and estimated delivery volume. Only after review can the system prepare multilingual alerts in Arabic, English, Turkish, Italian, and Hebrew.

Geographic delivery

The backend calculates targeting zones from an event location and severity using distance checks. Socket.IO keeps the analyst interface current, Sequelize persists the workflow in SQLite, and Web Push supports the notification path.

Prototype boundaries

The system is experimental. Its confidence model is heuristic, SQLite and application-level geographic filtering limit scale, and analyst authentication and authorization are not yet complete. Synthetic fallback evidence is useful for demonstrations but must never be confused with real corroboration in a production emergency workflow. The backend test suite passes 24 tests, and the frontend production build completes with a bundle-size warning that should be addressed through code splitting.