Debrief

Brief

Students can follow a lesson, answer the recap question, and still fail to explain or apply the idea later. Most learning tools test recall or ask whether something makes sense. Neither shows where a learner's mental model stops holding when they have to explain it in their own words.

Build

Debrief runs a four-step teach-back loop: Explain, Probe, Reveal, Repair. It maps an explanation into essential claims, finds the weakest one, asks a targeted follow-up, and shows the exact sentence where the reasoning broke before teaching that claim. The result is a claim-by-claim understanding map, never a numeric score. Curated lessons and an open-concept flow use the same loop.

Key decisions

• The model proposes, the reducer decides. Groq decomposes, evaluates, and writes questions, while a framework-free TypeScript state machine chooses the focus, gates each stage, enforces turn limits, and computes the verdict.
• Evidence quotes are constrained twice: only needs-attention evaluations can carry one, and every displayed quote must match the learner's original text after normalization. A missing match is discarded rather than paraphrased.
• There is no numeric score. A score would suggest a stable marking scheme that the product does not have, so the interface shows which claims are solid, unclear, untested, or need attention.
• Open-concept claims are derived from the topic and the submitted explanation in one validated Groq call. This replaced an earlier flow that generated protocol-stack claims from a topic name before it had seen what the learner wrote.

Outcome

Live at debrief.samuelyahaya.com and submitted to the SPEED August AI Challenge with a demo, gallery, and technical proof-of-work document. Live verification before recording exposed four real failures, including token truncation that had silently served fallback questions instead of model output. Devpost later removed the hackathon for a Community Guidelines or Terms of Service violation before judging, so there was no competition result. There are no user or adoption metrics to report.