System Architecture: End-to-End Flow of a Single Question

The full path

User → API Gateway → Classifier → Memory → LLM → YouTube → Response → Feedback → Eval (judge model).

Tech stack

• Edge — Cloudflare Workers + TanStack Start for the SSR app.
• Classifier — Claude Haiku via a dedicated worker.
• Memory — Postgres + pgvector for user profile and recall.
• LLM — GPT-4o-mini for answers, GPT-4o for hard cases.
• Video — YouTube MCP.
• Eval — nightly DAG, judge prompt on a larger model.

Multi-tenancy

Every record carries a tenant_id. Institutions get their own row-level-secured slice, custom system prompts, and per-tenant analytics.

Observability

Every request emits a trace with spans for each box above. p95 end-to-end is currently 2.8s.