JolSetu — the AI That Reads a 20-Taka Arsenic Test Kit (So Families Can Find Safe Water)

---

title: "JolSetu \u2014 the AI That Reads a 20-Taka Arsenic Test Kit (So Families Can Find Safe Water)" published: true description: "Computer vision reads a Gutzeit arsenic kit by photo, a QR-stamped live map tracks safe tubewells, and routing walks families to clean water \u2014 offline, in Bangla." tags: ai, bangladesh, showdev, opensource cover_image: https://rudra496.github.io/jolsetu/og-card.png

The problem, in numbers

• 20 million Bangladeshis drink water above the national arsenic standard (50 µg/L); 45 million above the WHO guideline — Flanagan et al., Bulletin of the WHO, 2012
• 24,000+ adult deaths every year are attributed to arsenic above 50 µg/L, plus up to 19,000 more in the 10–50 µg/L band — same source
• WHO called it "the largest poisoning of a population in history" (Smith et al., 2000)

Bangladesh ran massive national campaigns — 10M+ tubewells tested (BAMWSP 2000–2005, ARRMP 2021–23). But testing is once-in-decades, readings are human color-chart guesses, and when a well fails, families have no way to find the nearest verified safe one.

What I built

JolSetu (জলসেতু — "water bridge") attacks all three failures. Built as a university-track prototype for the Xylem Global Student Innovation Challenge 2026, water-quality problem statement.

Try it: rudra496.github.io/jolsetu

The four pillars

1. Test Kit AI — photograph the Gutzeit reaction pad next to the white reference strip; on-device color science (white-reference normalization + CIELAB matching) reads the chart and speaks the verdict in Bangla. The approach follows validated mobile-phone photometry of field kits (Haque et al., Sci Total Environ 2018).

Why it matters: on 943 real kit-vs-lab comparisons in the repo's dataset, human reading caught 99.3% of unsafe wells but wrongly condemned 37% of safe ones. Standardized AI reading attacks exactly that error.

2. QR-stamped live map — every check-in is HMAC-signed, GPS-geofenced (30 m pass/flag/reject at 2× radius), duplicate-checked per day, and reviewable. Trust is a design constraint, not a feature.

3. Predict & prioritize — a logistic risk model trained on real wells (AUC 0.67 on held-out data — disclosed, not inflated) flags where model and map disagree; the crowd tests there next. Active learning on real infrastructure.

4. Safe-well routing — distance, compass direction and walking time to the nearest verified safe wells. Works offline (service worker + cached tiles), in Bangla, with text-to-speech.

Built on real data, not a mockup

All 6,593 wells in the prototype are real measurements from the Araihazar upazila survey (Narayananj, Bangladesh), via CUAHSI HydroShare: DOI 10.4211/hs.8e1373d87419447c945625af13f0a2ea. The repo ships an EVIDENCE.md with a claim→source map, every DOI verified live, plus a model card.

Honest limitations

The colorimetry assumes the standard kit's chemistry and reasonable lighting; the risk model is deliberately simple and its AUC is published; the map's trust model depends on review participation. This is a prototype with a real evidence base — not a deployed system.

Links

• Live demo: rudra496.github.io/jolsetu
• Source: github.com/rudra496/jolsetu

If you work on water safety, ICT4D, or low-resource computer vision, I'd genuinely love your feedback — the issues tab is open.