Proof of concept · built for the Tonsser Lead Backend role
AI-assisted youth-player discovery, event-sourced end to end.
A scout asks for a report. An LLM analyses only the player's recorded stats, a guardrail rejects anything it can't prove, and the result is stored as an immutable event you can replay. This is the exact shape of Tonsser's problem — discovery, mobile-scale reads, AI as a first-class concern.
No sign-up. Realistic data is already loaded.
Architecture you can see
Commands → events → projections. Watch the live event stream rebuild read models in real time.
Open the event stream →AI with a guardrail
Every AI report is grounding-checked against recorded stats. Hallucinated numbers are rejected, not shown.
Generate a report →Evaluated, not vibes
A labelled eval suite scores grounding, schema validity, verdict match, latency and cost.
See the eval dashboard →