Ezra Learn

Inspiration

Many learning tools begin with where a child is expected to be. Ezra Learn begins with the evidence in front of the learner.

Faithful & True created Ezra Learn to explore a more patient approach: find the first missing foundation, change the teaching method when an explanation does not connect, and confirm understanding before moving forward.

What it does

In this focused prototype, fictional learner Maya completes a gentle mathematics warm-up. Ezra Learn recognizes repeated addition as a strength and identifies equal groups as the best next foundation before division facts.

During an $84 \div 6$ lesson, Maya can request another explanation. Ezra changes the representation from equal groups to a related multiplication fact. Maya then solves a new problem, $72 \div 6$, and explains why $6 \times 12 = 72$.

That transfer evidence helps update the demonstrated skill state and create a concise parent insight. The prototype also demonstrates homework-integrity redirection and a limited Trusted Grown-Up Bridge.

No real child information is used or required.

How we built it

Ezra Learn is a responsive web prototype built with React, TypeScript, the OpenAI JavaScript SDK, the Responses API, and Zod structured-output validation.

The tutor endpoint is designed for gpt-5.6-sol. GPT-5.6 interprets a learner message, identifies a likely misconception, chooses an allowed teaching strategy, and writes a brief child-facing response with structured learning evidence.

Deterministic application logic—not the model—controls navigation, transfer requirements, skill-state changes, homework-integrity redirection, and urgent-safety fallback routing. If a live response is unavailable, the experience remains usable through a clearly labeled deterministic demo mode.

Codex helped Faithful & True turn the broader Project Lighthouse specification into a focused Build Week prototype. It supported implementation, structured-output design, deterministic safeguards, automated testing, responsive refinement, deployment inspection, and submission documentation.

Challenges we ran into

The first challenge was making adaptation meaningful. Simply rewording an explanation was not enough. Ezra Learn needed to identify the missing foundation, visibly change the representation, and ask the learner to apply the idea to a new problem.

The second challenge was separating AI assistance from decisions that should remain predictable. We designed the model to recommend a teaching response while deterministic application logic controls lasting skill updates and safety routing.

We also needed the demonstration to remain reliable without misrepresenting a fallback as a live model response. Ezra Learn clearly labels which response source is being shown.

Accomplishments that we're proud of

We built a complete learning loop rather than a stand-alone chatbot demonstration: assessment, foundational-gap identification, teaching-method change, transfer problem, mastery update, parent insight, homework-integrity response, and a limited Trusted Grown-Up Bridge demonstration.

We are also proud that the prototype uses only fictional learner data, works across desktop and mobile layouts, includes a deterministic fallback, and passed its production build, lint checks, and five automated tests.

What we learned

We learned that adaptive learning is not simply generating more explanations. A useful system must understand which foundation is missing, select a genuinely different teaching representation, and look for new evidence of understanding.

We also learned that clearly limiting the model’s authority makes the experience more understandable and trustworthy. The model can support instruction while the application retains control of permanent state changes and safety boundaries.

What's next for Ezra Learn

The next technical step is to verify the live hosted GPT-5.6 response while preserving the reliable fallback.

Beyond Build Week, we would expand the prerequisite map beyond this focused division lesson, test additional teaching representations, improve accessibility and usability, and seek structured feedback from educators and caregivers.

Before using real learner information or conducting any real-world pilot, Ezra Learn would also need dedicated privacy, security, child-safety, and professional review. The Trusted Grown-Up Bridge would remain limited until it receives appropriate expert guidance and validation.

Built With

• codex
• gpt-5.6
• next.js
• node.js
• openai
• react
• responses
• tailwindcss
• typescript
• vite
• zod