Safety before generation
A deterministic safety layer evaluates critical signals before an AI response is composed.
02
Joy / Rehabify Health
AI-guided care access with safety outside the model
GeminiTypeScriptPostgresWhatsApp API
Overview
Joy is Rehabify Health’s WhatsApp-native care companion. Patients can describe what hurts in English, Yoruba, Igbo, or Hausa, move through intake and matching, book a verified physiotherapist, and continue recovery follow-up in a channel they already use.
From first symptom,
to continued recovery.
01 / Challenge
The conditions that shaped the system.
A useful architecture begins with the real constraint.
The experience needed to feel as simple as a conversation while coordinating sensitive health context, clinical red flags, identity, permissions, therapist matching, bookings, and payments. Safety-critical behavior could not depend on a generative model alone.
02 / System map
From input to outcome
01Patient on WhatsApp
→02Identity router
→03Safety rules
→04Gemini
→05Care context
→06Matching & booking
03 / Approach
How the build was organised.
Make the boundaries clear, then move.
- 01
Separated deterministic clinical red-flag checks from generative responses so urgent conditions follow explicit rules.
- 02
Built identity-aware routing and persistent context for conversations that continue across sessions and care stages.
- 03
Connected therapist matching, booking, payment, and follow-up actions to the same WhatsApp-native journey.
Decisions that carried the product.
Persistent context
Conversation state and patient context survive session boundaries so the assistant can continue coherently.
Channel-native workflow
Bookings and follow-up actions stay close to the WhatsApp experience patients already understand.
What changed after the work.
- ↗Unified intake, therapist matching, booking, payment, and follow-up workflows.
- ↗Created a multilingual assistant architecture with explicit clinical safety boundaries.
- ↗Established reusable identity and context patterns for long-running recovery conversations.
