CrisisWatchAI

CrisisWatch AI landing page for turning global disaster feeds into actionable intelligence.
Global crisis analytics dashboard showing events over time, severity distribution, and affected countries.
JKUAT Conduit station historical rainfall replay and inspection threshold workflow.
Real-time alert feed tracking critical and catastrophic crisis events globally.

Inspiration A campus facilities team gets a rainfall reading and asks a very practical question: should we check the drainage? Weather data is everywhere, but it rarely connects to an operational decision anyone can act on. We wanted to close that gap — turn station observations into a documented, reviewable inspection decision instead of another dashboard people glance at and ignore. What it does CrisisWatch AI replays a JKUAT Conduit weather export, estimates trailing 24-hour rainfall from the cumulative rainfall counter, and compares it against an adjustable inspection threshold. When a reading crosses the line, a proposed facilities user can save a historical drainage-inspection drill — with the source fingerprint, observation, threshold, and calculation evidence attached. Outcomes and follow-up notes append to history, and the whole thing exports as an evidence brief. A separate dashboard aggregates external crisis reports (GDACS, USGS, NASA FIRMS, ReliefWeb, WHO, ACLED, Open-Meteo) with a provider-health panel. It's a deterministic workflow — not a hype-driven flood predictor. How we built it

Backend: Python 3.12, FastAPI, SQLAlchemy + SQLite, pandas for the rainfall calculations, httpx for provider ingestion.
Frontend: Next.js, React, TypeScript, Tailwind, React-Leaflet for maps, Recharts for charts, TanStack Query for data fetching.
Data: The 2,943-observation Conduit export (Apr–May 2026) plus seven external crisis feeds with background refresh, dedup, and bad-row isolation.
ML: Experimental severity/text classifiers, forecasting and clustering — with fallbacks so the core workflow never depends on model availability.
Tests & CI: Unit suites for counter resets, gaps, window boundaries, dedup and source-hash validation; GitHub Actions CI. Challenges we ran into
Counter semantics: rainfall is inferred from increases in a cumulative counter — resets, invalid values, and long gaps all had to be detected or the estimate would silently mislead. Missing history intentionally withholds an assessment.
Provenance honesty: the CSV has no station ID, coordinates, or embedded proof of origin. We built source-hash fingerprinting and refused to claim organizer-approved eligibility or live API access we don't have.
Seven flaky external feeds: every provider behaves differently; we built per-adapter health reporting so failures are visible instead of silent. Accomplishments that we're proud of
A defensible, transparent rainfall calculation with an adjustable threshold and full evidence trail.
A complete drill workflow: assign owner + location, save evidence, append outcomes, export a brief.
Clean separation between deterministic station evidence and experimental ML — no overclaiming.
Engineering hygiene: XSS escaping, source-hash validation, duplicate prevention, and a passing test suite. What we learned
Honesty is a feature: stating what a system does not prove builds credibility with judges and users.
The hard part of climate-tech isn't modeling — it's messy, unverified field data and turning observations into decisions people can document and trust. What's next for CrisisWatchAI
Validate thresholds and response procedures with real facilities staff and run a pilot.
Verify CSV provenance and negotiate data rights with JKUAT Conduit.
Add user-level authorization, migrations, and durable scheduling for shared deployment.
Connect inspection outcomes back into the dashboard as feedback.