Lantern

Inspiration

AI tutors are very good at producing answers. In education, that can short-circuit the most valuable part of learning: explaining why an approach made sense. Lantern starts from the learner's own words and treats an incorrect response as evidence of a current conceptual model.

What it does

The Learner studio analyzes a problem and the learner's reasoning. It returns a likely misconception, a three-part concept map, three Socratic questions, a micro-exercise, and one teacher move. The Class map demonstrates how multiple reasoning traces can be grouped into misconception clusters and converted into a focused lesson recommendation.

How we built it

Lantern is a React and Express application. The server uses the OpenAI Responses API with GPT-5.6. A Zod schema and Structured Outputs keep the analysis reliable for the product interface. The prompt prevents direct answer disclosure and requests age-neutral, non-clinical instructional language. A deterministic demo mode keeps the complete experience testable without exposing an API key.

How Codex helped

Codex was used for the majority of the implementation. It translated the product concept into a working full-stack prototype, implemented the GPT-5.6 schema and prompt, created responsive student and teacher views, added tests, and verified the complete experience at desktop and mobile sizes.

Challenges we ran into

The central product challenge was avoiding an answer machine disguised as a tutor. The output contract was designed around evidence, conceptual bridges, questions, and small exercises. The second challenge was keeping the demo reliable while protecting credentials, which led to a clearly labeled no-key demo mode.

Accomplishments that we're proud of

• A complete reasoning-to-intervention flow that is easy to demonstrate in under three minutes
• Stable GPT-5.6 outputs through a structured schema
• A class-level view that connects individual diagnosis to teacher action
• A responsive, accessible interface with loading, empty, error, and demo states
• Local tests and a production build with zero known package vulnerabilities

What we learned

Structured AI output is most useful when the product team decides what the model should not do. For Lantern, refusing to reveal the answer shaped the schema, prompt, and interface more than any individual feature.

What's next for Lantern

The next version would accept image-based student work, let teachers confirm or revise a diagnosis, and build a privacy-preserving misconception timeline across assignments. Any school deployment would require stronger student-data controls, consent policies, and educator evaluation.

Built With

• codex
• express.js
• gpt-5.6
• node.js
• openai-responses-api
• react
• vite
• zod