Mali

Inspiration

AI learning tools rely on open-ended chat, which can feel unstructured and leave you unsure if you've actually mastered a topic. I built Mali to combine the natural, adaptive dialogue of an LLM tutor with a structured curriculum

What it does

Mali takes any topic you want to learn and automatically generates a step-by-step skill curriculum. Powered by GPT-5.6, it guides you through interactive lessons one skill at a time, evaluating your understanding along the way and tracking progress deterministically

How we built it

I composed formal specification drafts first, then used OpenAI Codex as our primary developer to implement the core engine, web app, and property-based test suites. Inside the application, GPT-5.6 handles two key jobs: generating structured curricula with computable answer keys, and conducting real-time, context-aware tutoring sessions.

Challenges we ran into

The biggest challenge was bridging open-ended LLM dialogue with strict system verification.

Accomplishments that we're proud of

I am proud of my spec-driven workflow with Codex, which enabled us to build a heavily tested codebase quickly and cleanly.

What we learned

I learnt about knowledge state graphs, where by representing curricula as partially ordered skill graphs and tracking progress through valid knowledge states (downsets), it can strictly be guaranteed that prerequisites are always satisfied.

What's next for Mali

I plan to expand evaluation to support qualitative subjects, add multimodal support so learners can upload photos of handwritten work, and introduce spaced-repetition schedules to ensure long-term skill retention.

Built With

• codex
• gpt
• openai
• python