TraceLearn

See the thinking, not just the answer.

Comment

Inspiration

Teachers often see a final wrong answer while the student's actual reasoning remains hidden in crossed-out lines, skipped steps, or a single misapplied rule. Existing AI tutoring tools often jump directly to the correct solution, erasing the most valuable teaching moment: understanding why the student's answer felt reasonable.

TraceLearn asks a different question: what if AI made student thinking inspectable without replacing teacher judgment?

What it does

Teachers paste a problem and a student's work. TraceLearn:

  1. reconstructs a concise sequence of observable reasoning moves;
  2. marks the earliest likely misconception;
  3. distinguishes downstream consequences from additional errors;
  4. shows evidence and confidence for its hypothesis;
  5. proposes one Socratic question that can reveal or repair the mental model;
  6. offers alternate questions and a concise teacher note.

Every result visibly requires teacher review. TraceLearn does not diagnose students or assign traits—it proposes a testable instructional hypothesis.

How we built it

The frontend is a responsive React + Vite workspace designed around an open reasoning canvas rather than a generic dashboard. The live API uses GPT-5.6 through the OpenAI Responses API. Zod Structured Outputs enforce a stable contract for reasoning nodes, statuses, misconception index, evidence, confidence, coaching questions, and teacher notes.

The server sends student work with store: false, derives a privacy-preserving safety identifier from a local anonymous session ID, and uses medium reasoning effort. A deterministic demo mode keeps the complete judging flow runnable without credentials or external setup.

Codex was the primary build environment for challenge research, product framing, visual direction, component architecture, API implementation, tests, documentation, and browser-based desktop/mobile validation.

Challenges

  • Showing uncertainty without weakening usefulness. Confidence, evidence, and teacher review are part of the product surface rather than hidden caveats.
  • Separating cause from consequence. Later algebra can remain coherent after one early mistake, so the schema distinguishes misconception from consequence nodes.
  • Making the demo reliable. Judges can use the full deterministic flow immediately, while the same UI switches to live GPT-5.6 when a key is present.
  • Avoiding another card dashboard. The main experience is an annotated reasoning path with teacher-margin notes.

Accomplishments

  • A complete teacher workflow rather than a prompt wrapper.
  • An inspectable visual explanation of the model's hypothesis.
  • Strict structured output with validation and refusal/error handling.
  • Privacy-conscious defaults and a clear human-review boundary.
  • A no-credential judging path, responsive layout, automated tests, and production server.

What we learned

The most useful educational AI output is often not an answer—it is a well-framed next observation or question. Structured Outputs also strengthened the product design: once the reasoning trace became a contract rather than prose, we could design evidence rows, confidence treatment, and multiple coaching moves around it.

What's next

  • Image input for handwritten student work with teacher-controlled transcription review.
  • Longitudinal misconception patterns without turning them into permanent student labels.
  • Teacher-authored rubrics and curriculum-specific diagnostic policies.
  • Classroom-safe sharing, multilingual coaching questions, and expert-teacher evals.

Try it

The public demo requires no login or API key. Select Analyze reasoning to run the deterministic judging flow. The README contains local setup, production deployment, testing, architecture, and detailed notes on how Codex and GPT-5.6 were used.

Built With

Share this project:

Updates