A role-based platform that checks MPPCB ghat data against CPCB criteria in code, ranks ghats for review and drafts action briefs that a person reviews.
The problem
Bhopal's immersion ghats are tested by MPPCB around Ganesh and Durga visarjan, but the results sit in PDFs and spreadsheets. Officers can't quickly see which ghat needs attention, how this season compares with last year, or whether a reading actually breaches the CPCB limits. Ghat Sentinel turns the 2023-2025 MPPCB data into a reviewable evidence trail. It shows which ghats breach CPCB criteria, ranks them for review, and drafts action briefs. A person always reviews the result, and every number traces back to its source file, row and PDF page.
How it works
Ghat Sentinel is a role-based web platform. Officers and workers log in, and community members get a plain-language guide.
- Deterministic CPCB criteria engine. Class B limits apply only to pH, dissolved oxygen and BOD. Other parameters are marked "not comparable". Missing or BDL values are never treated as zero.
- Review-priority score per ghat with a visible factor breakdown (40/20/25/15). It reports "insufficient data" when there are fewer than 6 usable readings.
- Season heatmap and trends across Ganesh and Durga visarjan, 2023-2025. The data has no exact dates and no coordinates, so we show seasons and phases, not a map.
- Honest forecasting: the Oct 20 outlook refuses to predict because there are only 3 Durga seasons. It returns a historical comparison instead.
- Verification cases with a status workflow, a downloadable evidence report (JSON/CSV), and Bhopal vs statewide benchmarking where the data is comparable.
- Claude-written briefs and citizen Q&A, strictly validated: any cited reading, number or link that isn't in the evidence is rejected and a template is used instead.
- Officers can send briefs by email (SMTP) and WhatsApp (Cloud API) straight from the app, with a message log and rate limits.
- Roles: officer, worker, community, developer-only system status, and a presenter demo account.
How Claude was used
Product: the Claude API (claude-haiku-4-5) writes the action briefs and answers citizen questions. It only gets a pre-computed evidence pack. Output is validated so it can't cite readings, numbers or URLs outside that pack, and if validation fails the app falls back to deterministic template text. The CPCB comparisons, scores and forecasts are all computed in code, not by AI. Development: we used Claude Code to review and complete the backend, build and integrate the frontend, add authentication and roles, write the email and WhatsApp integration, write tests, and set up Docker and the Render/Neon/Vercel deployment.
Team
AquaSentinel