An AI marker that grades work and explains itself
In production
Built for an education client. Submitted work is queued to an asynchronous evaluation worker that scores it against a rubric and returns structured, itemised feedback rather than a bare number. Every result is reviewable, and a human can override any score — the model proposes, a person decides.
The queue is the point: marking runs independently of the app, so a slow or failed evaluation never blocks a child or a parent, and a failed job retries with its error recorded.
- FastAPI
- Async worker queue
- LLM evaluation
- Postgres
- Rubric scoring
- Human override