DiaCare AI is built by a single-digit team. Here is how it works.
The model
DiaCare runs on Arka Helix 1 — the same wrapper model that powers Arka AI. Helix 1 is a careful composition of a high-quality open-weights foundation, Vyoma's system prompt layer, and a memory + safety overlay.
For DiaCare, the system prompt is tuned heavily toward diabetes education, South Asian context, Indian foods and units, and conservative safety. The prompt is ~6,500 characters and includes explicit refusal patterns for diagnosis and prescription.
The backend
A single Cloudflare Worker. Receives the chat messages array, adds the model selection (helix-1), forwards to the inference layer, streams the response back. Worker-level rate limiting, abuse detection, and abort handling. No database — the Worker is stateless.
Why Cloudflare Workers? Low latency from anywhere (especially India), generous free tier, edge-native, easy to evolve. No server to manage.
Health profile injection
User health profile lives in browser localStorage — not on our servers. On every chat call, the profile is appended to the system prompt under a clear "USER HEALTH PROFILE" tag:
- Basic info (name, age, type of diabetes if applicable)
- Recent blood sugar logs (if entered)
- Symptoms recently reported
- Medications (if entered)
This lets Helix 1 personalise every response without us ever holding sensitive health data on a server.
Safety design
Three layers:
- Prompt-level refusals. Helix 1 is instructed to refuse diagnosis, prescription, and emergency advice, and to escalate to "call emergency services" for DKA, severe hypo, chest pain, etc.
- UI-level reminders. The disclaimer "DiaCare AI provides educational info only" appears at the bottom of every screen, linking to the full safety guide.
- Tone-level honesty. When Helix 1 is uncertain or the user's question is outside scope, it says so plainly instead of guessing.
What is next
Voice-first mode for low-literacy users. CGM data import. Doctor share — one-tap export of recent logs to take to your next appointment. Hindi and Tamil prompt variants tuned for those languages, not just translated. Read more in our DiaCare documentation.