We will be undergoing planned maintenance on Oct 7th 6:00AM UTC / Oct 7th 2:00AM ET

JolSetu

Explainable AI reads a $1 arsenic test kit; humans confirm every result; the community gets a verified map of safe water.

Comment

Track

Track 3 — AI-Supported Assessment (with strong Citizen Science UX elements from Track 1).

Inspiration

In Bangladesh, 61 of 64 districts have groundwater arsenic above the WHO limit, and millions of tubewells are untested. Testing exists — a $1 Gutzeit kit makes it possible anywhere — but people cannot read the result reliably, the data dies in paper notebooks, and families keep drinking from unsafe wells. JolSetu (জলসেতু, "water bridge") was built to close that loop for real communities in Araihazar, and this hackathon is our chance to harden it with the OneAquaHealth mission: healthy waters, healthy ecosystems, healthy communities.

What it does

JolSetu is an offline-first PWA that turns any citizen into a reliable water tester:

  1. AI-assisted kit reading (Track 3 core): the phone camera images a $1 Gutzeit test kit; a deterministic computer-vision pipeline (CIELAB colorimetry with a white reference card, per-channel analysis) estimates the arsenic level — fully explainable: users see the exact color channels and reference correction behind every number.
  2. Human-in-the-loop by design: the AI never has the final word. Every reading must be confirmed or corrected by the tester; borderline results trigger a re-test workflow; community-confirmed readings enter an active-learning queue that improves future estimates.
  3. From reading to action: HMAC-signed QR + geofenced check-ins register each well; a live map of 6,593 real Araihazar wells (public HydroShare dataset) powers safe-well routing; a sklearn risk model (AUC 0.672 on a real 80/20 split) flags priority wells for the DPHE (public water authority) dashboard.
  4. For everyone: Bangla + English UI with spoken (TTS) results, designed for low-literacy users and offline rural use.

How we built it

Vanilla-JS offline-first PWA (service worker, installable, works in airplane mode), pure client-side computer vision (CIELAB color distance + white-reference normalization — no server, no API keys), Leaflet mapping over 6,593 real geo-located wells, scikit-learn risk model trained on public data, Web Speech TTS in Bangla/English, HMAC-QR anti-fraud check-ins. Public repo, public dataset, public dashboard.

Challenges we ran into

Camera sensors lie about color: we solved it with a printed white reference card + per-channel normalization instead of a black-box model, keeping the math explainable to a regulator. Field connectivity is zero: everything runs offline and syncs later. And trust: water authorities will not act on unsupervised AI — so human confirmation is not a UX extra, it is the architecture.

Accomplishments that we're proud of

A complete, working, deployable pipeline: kit photo → explainable AI estimate → human confirmation → verified public map → safe-well routing → authority dashboard — running on a $0 infrastructure budget and a $1 test kit.

What we learned

Responsible AI in citizen science is 10% model and 90% workflow: validation checks, uncertainty language, and the human confirmation loop are what make citizen data trustworthy enough to act on.

What's next

Adapting the same pipeline to OneAquaHealth stream-assessment kits (nitrate/phosphate colorimetric tests share the same CIELAB backbone), school club modules, and DPHE-scale open dashboards.

Built With

Share this project:

Updates

Submission history