Inspiration
Millions of people rely on shared water points like wells and pumps. When one breaks, the people who can fix it often never hear about it. A caller may not know a water point's name, and many have no smartphones or data. A phone call works for almost everyone. We wanted anyone to be able to report a broken water point with a phone call, and NGO staff to be able to pass those reports to trusted repair partners.
What it does
A caller phones the number and talks to an AI voice assistant. It asks where the water point is and what's wrong, including water points the system doesn't know yet. The server saves the report, matches or creates the water point, and scores the complaint. NGO staff sign in, track complaints, pass them to partners and mark them fixed.
Nothing is dropped. If a call is unclear, the location is unusable, or several water points fit, the report goes to a Needs review page for a human to resolve.
We store no caller names or phone numbers. We do not offer water-safety advice and do not guarantee repairs. A human confirms every action.
How we built it
It is a Next.js web app with PostgreSQL + Drizzle and Better Auth for sign in, served through a Cloudflare tunnel. Vapi runs the phone line.
The webhook is a small pipeline: check the secret, validate and scrub the payload, and save it. If matching or scoring fails after the first save, the report is still kept.
We wrote the spec, the data contracts and the rules first, in docs/. Six AI agents (Orchestrator, Backend, Frontend, Security, Phone/AI Prompter, UX) worked in three rounds of Plan, Build, and Review. Humans approved every merge through pull requests with CI.
Challenges we ran into
- The assistant could pick up a person's name in the location or language fields. We tightened the prompt wording, scrubbed numbers in code, and documented the remaining risk.
- In one test, the assistant picked up other people's voices and got confused.
- Missing or unknown fields must never crash the endpoint, so we built it to always keep the report.
- A safety guard meant for tests would have blocked production.
- Tests with hard-coded dates began failing once the real clock passed them, and we fixed that by freezing time in the tests.
- We had to coordinate many agents and people without stepping on each other: one task, one branch, one small pull request.
Accomplishments that we're proud of
- A spoken report appears in the staff app with no caller name or number.
- Every odd call is kept and explained on the Needs review page.
- A clear set of red-line rules (privacy, human confirmation, server-side permissions, no overwriting history) that the agents and the tests actually follow.
- We shipped a working, documented, tested system as a small team of mostly first-time developers.
What we learned
- Voice AI needs strict rules and fallbacks, because real callers are noisy, unclear and unpredictable.
- Writing contracts and rules before code kept six agents and multiple people consistent.
- Tests should control the clock.
- "Never drop a report" is a design rule, not a feature. It shaped the database, the pages and the error handling.
What's next for Well Watch
- Switch the 15-second auto-refresh to live updates
- Support more languages and regions, and let more NGOs and partner organizations onboard
- Smoother handoff to repair partners, with status updates back to the NGO
- Better matching of unknown water points with a human in the loop
- A stronger privacy check on location and language text, and a review of the open decisions in the docs
Built With
- betterauth
- claude
- cloudflare
- codex
- docker
- drizzle
- github
- javascript
- next.js
- orm
- postgresql
- sql
- typescript
- vapi