AI Health Chat
Multilingual health-information and triage-support assistant on web and WhatsApp
1st place, Power Learn Project Class of 2024 graduate hackathon (December 2024)
A public-beta assistant that gives educational health guidance in five languages, on an installable web app and on WhatsApp. It keeps track of an ongoing case across turns, escalates consistently when symptoms need professional care, and can produce a SOAP-style summary the user chooses to share with a clinician.
At a glance
- Role
- Founder and lead engineer
- Status
- Public beta
- Type of work
- Independent product
- Years
- 2024 – present
- Categories
- Applied AI & Agents · Web & SaaS · Civic & Social Impact
- Source code
- Source private
Technology
Problem
Many people look up health concerns on a phone first, often in a language other than English, and need to know when a symptom calls for professional care. General-purpose chatbots rarely keep a consistent picture of the case or apply the same safety rules in every language.
Who it’s for
People looking for general health information in English, Kiswahili, French, Spanish or Dutch, and the clinicians they later see, who can receive a structured summary the user chooses to share.
What I built
- Web PWA and Meta WhatsApp integration served by the same assistant logic.
- Medium and High detail response modes with consistent safety escalation and a healthcare disclaimer.
- Conversation continuity and triage routing, with a multi-turn case state so follow-up questions refer back to the right case.
- Image analysis with Mistral/Pixtral, deliberately conservative and limited to first-aid level guidance.
- Biometric tracking: manual entry plus opted-in extraction from chat, with validation and trend charts.
- SOAP-style consultation reports built only from measured or logged vitals, with provenance for each value, and privacy-minded share links.
- Provider fallback across Groq, OpenRouter and Mistral so one provider outage does not take the assistant down.
Architecture
- ChannelsInstallable web app (PWA) and a Meta WhatsApp webhook with request verification.
- OrchestrationServer-side routing, case state and detail modes; provider keys never reach the browser.
- ModelsGroq and OpenRouter text models with fallback; Mistral/Pixtral for images.
- DataSupabase PostgreSQL with Row Level Security, opt-in consent and retention purging.
- OutputsGuidance with disclaimers and escalation; SOAP reports with vitals provenance; share links.
Engineering challenges
Keeping a long conversation coherent
A structured case state carries symptoms and context between turns, so a follow-up like "it is worse today" is interpreted against the right case instead of starting over.
Applying the same safety rules in five languages
Escalation guardrails and disclaimers apply regardless of language or detail mode, and the test suite includes checks across all five languages.
Reports a clinician can trust
SOAP summaries only use vitals that were measured or logged and record where each value came from, rather than letting the model infer numbers.
Quality and trust
- Automated test suite, which the project README reports as 144 passing tests, including five-language checks.
- Supabase Row Level Security, server-side API keys, webhook verification and rate limits.
- Opt-in consent, retention purging and a visible healthcare disclaimer.
Results
| Measure | Value | Evidence |
|---|---|---|
| Hackathon | 1st place, PLP Class of 2024 graduate hackathon | Owner-reported, Dec 2024 |
| Automated tests | 144 passing | README-reported |
| Languages | 5 | Product feature |
Limitations
- Provides educational guidance and triage support. It is not a diagnostic tool, has not been clinically validated, and does not prescribe or replace a clinician.
- Image analysis is intentionally conservative and limited.
- Public beta: features and behaviour are still changing.