GhatWatch

The problem: Every Ganesh and Durga festival, MPPCB tests the water at Bhopal’s 7 immersion ghats three times (before, during and one week after). The results sit in 90-120 page PDFs that nobody lines up against the legal limits or compares year to year. The BMC officer who plans Durga visarjan (20 October 2026) decides where to send cleaning, aeration and extra testing, and today that decision rests on memory.

What we built: GhatWatch gives that officer one dashboard that answers “where do I send what, before 20 October?”

Ranked ghats: ranked by the share of pH, DO and BOD readings outside the CPCB class B limits. The ranking is rule-based, so the same data always gives the same order. Traceable numbers: every number links to the report and PDF page it came from. Actions: each ghat gets a suggested action (aeration, cleaning or testing) and a plain-language reason. Missing data: Bairagarh and Malikhedi have no Durga 2025 table, and they are among the worst ghats on dissolved oxygen. The tool shows them as “data missing, test now”, never as clean.

SOLUTION GhatWatch is a 3-tab dashboard that answers one question for the BMC officer: "Where do I send what, before 20 October?"

Tab 1 - Analysis

• 10 KPI cards: days to Durga, Tier 1 ghats, top-ranked ghat, DO/BOD/pH limit failures, zero-DO readings, data completeness (88%), pre-immersion BOD failures, ghats with no Durga 2025 table
• Rule-based tiered ranking of the 7 ghats by share of pH, DO and BOD readings outside CPCB class B limits, with fail-rate bars and confidence labels
• 4 charts: fail rate by parameter, coverage heatmap (missing data shown hatched grey), DO dot plot against the 5 mg/l limit, pre/during/after phases
• Durga / Ganesh / All toggle

Tab 2 - Actions & Insights

• Summary card and ghat-wise action cards (aeration, cleaning, testing) in priority order
• Rule-based alerts (critical, warning, data) and insights
• Ready-to-forward English brief with a copy button

Tab 3 - Map (approximate)

• Bhopal city map with a colour-coded circle around each ghat (red, orange or amber for tier, hatched grey for no data)
• Popups with fail rate, action and an Evidence button
• Warning strip for ghats that could not be placed
• Legend states it shows reading status, not pollution spread

Across the tool

• Evidence drawer: every number links to its report and PDF page
• Modern UI with hover effects, count-up numbers and animated charts
• Works offline from a double-click; no LLM at runtime, no API key

HOW WE USED CLAUDE AND CLAUDE CODE

Claude (chat)

• Explained the problem, corrected our assumptions (ranking must be rule-based, missing data is never "clean") and compared 3 scoring options. We picked exceedance rate because it is the easiest to trace to readings and CPCB limits.
• Designed the dashboard (KPIs, tiers, alerts, 3 tabs) and wrote our Claude Code prompts.
• Caught problems: an API key in a file about to go public, and blocked map tiles.

Claude Code

• Profiled the CSVs and PDFs, reproduced the problem statement's numbers (42/116, 100/112, 4 zero-DO readings) and spot-checked rows against PDF pages.
• An orchestrator wrote the data contract, then ran Scoring, Frontend and Brief agents in parallel, followed by a read-only Verifier (9/9 tests passed) and a Docs agent.
• Iterated on the redesign, the map and the styling, checking its own work with screenshots.