Skip to main content

AI Architecture

AI is designed to run as an isolated service (Python/FastAPI) behind explicit contracts, provider-agnostic via LiteLLM. It assists; it never decides medical outcomes.

Nothing on this page is running. apps/ai is a FastAPI application whose only routes are /healthz and /readyz — no business endpoint, no capability, no model call. It is not in infrastructure/coolify/docker-compose.prod.yml, so nothing deploys it. There is also an aiassist module inside the Go API; treat its status as separate from this page and check ROADMAP.md for it rather than assuming either way.

Everything below is therefore a target state and a set of constraints on it, not a description of behaviour you can call.

Guardrails (non-negotiable, for when it is built)​

  • No autonomous diagnosis or prescription.
  • Mandatory disclaimer on any health-adjacent output.
  • Human validation for critical veterinary outputs.
  • Prompt & model versioning; full traceability (input hash, model, latency).
  • Data minimization — callers pass only what a capability needs.
  • Country-specific activation checked by the core API.

Planned capabilities​

Owner assistant, appointment preparation, medical-document summarization, report explanation, translation, lost-&-found matching, image similarity, content moderation, fraud-detection assistance, quiz generation, professional note drafting, recommendations.

Status: a service skeleton with health routes only, not deployed. No capability listed above exists. The guardrails are recorded now precisely so they constrain the work when it starts, rather than being retrofitted to it.