My World — Private Family Learning Coach

A privacy-first family learning center that uses GPT-5.6 to turn local progress into accurate, actionable next-step learning plans.

Comment

Inspiration

Families often own a valuable learning library, but the files are scattered across videos and documents. A child still needs reliable answers to two questions: “Where did I stop?” and “What should I do next?” Uploading a private family library to a public service also creates unnecessary privacy risk.

What it does

My World is a private, child-friendly learning center. It keeps the original course library read-only on a family NAS and stores each learner’s progress locally. The home screen compares recent activity across courses and resumes the exact learner at the correct lesson. When a lesson is complete, it advances to the next unfinished lesson in that course.

On demand, GPT-5.6 turns a small, allowlisted set of course and progress metadata into a structured plan containing three next steps, a 10–45 minute time recommendation, review topics, exactly three reflection questions, and a short factual parent summary.

How we built it

The app uses React, TypeScript, Vite, Node.js, SQLite, and FFmpeg. The server calls the OpenAI Responses API with gpt-5.6-sol, low reasoning effort, store: false, and a strict JSON Schema. Codex was used to design and implement the core Build Week functionality, tests, privacy boundary, judge mode, and submission assets.

Only allowlisted metadata is serialized. Names, member IDs, file names, NAS paths, media, and session tokens are excluded. The browser never receives the OpenAI API key.

Privacy-safe judge mode

The reproducible demo uses anonymous learners, an isolated SQLite database, a small sample catalog, seeded progress, and synthetic videos. It never reads the private NAS or real family data. The demo intentionally shows that robotics resumes even when math is featured, because robotics has the learner’s latest progress timestamp.

Challenges

The hardest parts were making resume behavior correct across courses, enforcing a strict structured-output contract, and proving sensitive data never enters the model request. We also separated the public judging environment from private runtime data instead of hiding private data behind a visual toggle.

Accomplishments

  • A real GPT-5.6 Responses API request returned HTTP 200 and passed schema validation.
  • 23 automated tests pass.
  • The production build completes.
  • All 19 courses and 2,142 asset references pass catalog verification.
  • npm audit reports zero known vulnerabilities.
  • Failures are explicit; the app never substitutes a fake AI response.

What we learned

Deterministic progress logic and generative planning work best as separate layers. Local state decides where learning resumes; GPT-5.6 explains the most useful next steps. Privacy is strongest when enforced at the serialization boundary and covered by regression tests.

What’s next

Next steps include a parent authentication gate, richer learner profiles, more educational planning modes, and a public deployment of the anonymous judge environment.

Built With

Share this project:

Updates