Kitab

Inspiration

Most books are static: they can contain excellent material, but they do not know what a learner already understands or what they should study next. We wanted to turn any book into a personalized tutor—one that maps its ideas, diagnoses knowledge gaps, and guides learners through a prerequisite-safe path.

What it does

Kitab transforms any book into an interactive learning system.

For the demo, we used the openly available Illustrative Mathematics high-school curriculum to demonstrate the complete experience. However, user can upload any book that they have access to, in pdf or url. Kitab then parses the book and builds a knowledge graph- converting the material into atomic, measurable concepts and connecting them through prerequisite relationships.

The learner can walk through this knowledge graph, and then take a short adaptive diagnostic quiz. Each question is selected to gather useful evidence about the learner's current understanding. Afterward, Kitab provides a progress report with correct and incorrect responses, question-level feedback, and mastery signals visible directly on the knowledge graph.

Based on that evidence, Kitab creates a customized learning journey: the next useful skill, focused micro-lessons, clean mathematical notation, worked examples, source-grounded explanations, and practice.

How we built it

Kitab is packaged as a local Codex plugin with skills, an MCP learning engine, a web tutor, and persistent learner state.

The system includes:

• A book-to-knowledge-graph pipeline that extracts concepts, evidence, skills, and prerequisite relationships.
• A local MCP server that manages diagnostics, scoring, recommendations, mastery state, and lesson progression.
• A web interface for browsing the graph, taking diagnostics, reviewing progress, and completing lessons.
• Private scoring contracts, rubrics, source evidence, and validation rules for diagnostic items.
• A model-authority workflow using GPT-5.6: one model proposes material, another independently checks it, and higher-reasoning review adjudicates disagreements and audits release.

Codex was used throughout development and at runtime to orchestrate the local plugin workflow, including launching the tutor, coordinating MCP actions, and managing adaptive learning flows.

Challenges we ran into

A major challenge was making the experience genuinely adaptive. We needed diagnostic questions to be useful, fast to score, and accurate to show learners.

We also had to make a complex workflow feel simple: a learner should move smoothly from knowledge graph to diagnostic to report to micro-lessons, without waiting on unnecessary model calls.

Finally, turning source material into a trustworthy graph required careful handling of evidence, prerequisite relationships, validation, and unresolved concepts.

Accomplishments that we're proud of

We built more than a chatbot around a book. Kitab delivers an end-to-end personalized learning loop grounded in the material itself.

We are especially proud of:

• An interactive knowledge graph that learners can revisit at any point.
• Adaptive diagnostics with question-level progress reporting.
• A custom learning path with completed, current, and upcoming lessons.
• Local learner state and scoring workflows.
• Evidence-grounded micro-lessons and practice rather than generic AI responses.
• A reusable Codex-plugin architecture that can support new books and subjects.

What we learned

We learned that personalization is not just about generating explanations. It depends on knowing what a learner understands, what prerequisites are missing, and why a particular next step is useful.

We also learned the value of separating model roles. Proposal, independent checking, adjudication, and release auditing make the system more reliable than relying on a single model response.

Most importantly, we learned that a strong learning experience needs both intelligent backend logic and thoughtful UX. Learners should never need to understand the graph or scoring system before benefiting from it.

What's next for Kitab

Next, we want to support more uploaded books and subjects beyond mathematics, strengthen the ingestion pipeline, and improve the automated generation and validation of knowledge graphs.

We also plan to expand lesson formats, add richer learner analytics, support educator views, and make the generated tutors easier for publishers and open-course creators to deploy.

Kitab’s long-term goal is simple: turn any evidence-grounded book into a tutor that knows what a learner should learn next—and can explain why.

Built With

• codex