Insight AI

A Socratic AI coach that never gives you the answer — it asks the right question so you discover it yourself, then tests what you learned to make it stick.

Comment

Inspiration

We kept noticing the same problem: tools like ChatGPT and Gemini hand you the answer instantly, but that shortcut quietly kills learning. You get the fix, never the mental model that makes it stick. We wanted an AI that refuses to just tell you — one that acts like a real tutor, asking the right question at the right moment so the insight is yours.

What it does

Insight AI is a Socratic learning coach. Instead of giving answers, it guides you to them through micro-questions and analogies. Two things make it more than a chatbot: an active-recall "Test Me" engine that tracks the concepts you've worked on and auto-generates short quizzes with instant feedback, and cross-session memory that remembers where you struggled and picks the thread back up in your next session. When you're stuck, hint controls ("give me a hint", "nudge me closer") move you forward without ever handing over the answer.

How we built it

The product is operated by Google Gemini. Every Socratic turn, concept-tracking pass, and quiz generation/evaluation runs as an autonomous Gemini call, logged as an AI operation — the model drives the core learning loop, not a human. We designed the interface with Stitch, Google's Gemini-powered design tool, iterating on the UI through prompts and exporting the design as a spec. We used Claude Code as our development assistant to build against that spec while keeping a disciplined, modular codebase. The backend runs on Google Cloud Run (Express + TypeScript, JWT auth over httpOnly cookies), the frontend on Vercel (Next.js 14 + Tailwind), with data in PostgreSQL hosted on Supabase.

Challenges we ran into

Getting cross-site auth right between a Vercel frontend and a Cloud Run backend took real work — cookies, CORS, and SameSite settings all had to line up. Mid-build, Google retired the Gemini model we'd pinned, and our calls started returning 404s; we moved to a floating model alias so a future retirement can't break the demo. We also designed a two-layer free/paid quota (sessions per month + messages per session) that had to enforce cleanly in production without breaking the Socratic flow — and had to make sure hint controls never quietly leaked the answer.

Accomplishments that we're proud of

A Socratic engine that genuinely holds the line — it guides without revealing, even under "just tell me" pressure. Cross-session memory that recognizes when you return to a weak concept versus starting something new. And a real, deployed product with live auth, quotas, and a working payment path, not just a demo script.

What we learned

Prompt discipline matters as much as model choice — keeping Gemini in a consistent teaching posture across chat, quizzes, and memory was harder than any single feature. We also learned to design for model volatility: pinning a version felt safe until it got retired under us.

What's next for Insight AI

Sharper progress tracking per concept, richer active-recall formats, and expanding cross-session memory into a longer-term learning profile. On the product side, opening the paid tier to real learners and using their feedback to tune the Socratic prompts.

Built With

Share this project:

Updates