tutorbook

Inspiration

Students don’t need another chatbot that forgets their PDF. They need a Study OS — a notebook that turns course material into a durable wiki and Study Map, then tutors with Evidence you can open and verify.

TutorBook exists because generic AI chat invents citations and LMS tools store files without teaching. We wanted provenance, structure, and continuity across study sessions.

What it does

TutorBook (tutorbook.me) is a personal Study OS:

Start from a Published Study Template (or ingest sources)
Get a source-grounded Source Wiki + Study Map
Learn with a tutor that cites Evidence from the material
Compound progress — what to study next stays oriented across sessions

Fastest path for judges: sign in → consent → clone Algebra Foundations → Study Map → Source Wiki → ask the tutor → open Evidence.

How we built it

TypeScript monorepo: React/Vite learner UI, Fastify API, workers, Postgres/Neo4j, Pi tutor runtime.

Built with Codex + GPT-5.6 Sol:

We burned days and millions of tokens trying to design wiki reconciliation — how sources become a durable, evidence-grounded, atomically published wiki. It kept failing.

GPT-5.6 Sol fixed it in almost one cycle:

Sol high — ideation and planning (asked a few clarifying questions)
Sol light — end-to-end implementation

That became the bounded Wiki Reconciler (ADR-0040): durable stages, Evidence-grounded pages, atomic snapshot publish, deterministic Study Plans.

Then the same workflow built our evaluation system: gather data from many sources → clean/structure → eval sets → sandbox run pipeline → operator dashboard (/admin/eval). That stack is still landing locally; the architecture is real.

Challenges we ran into

Designing wiki reconciliation without queue soup (initial/rolling builds, parallel readiness flags) — the hard problem Sol finally cracked
Keeping learner-visible knowledge atomic (publish a full snapshot or nothing)
Making Education quality measurable: corpus → suites → isolated sandbox → dashboard, without writing false passes
Packaging a judge-ready demo without live ingestion risk (template-first Algebra path)

Accomplishments that we're proud of

A Study OS learners can try at tutorbook.me, not a slideware mock
Wiki Reconciler architecture after days of failed attempts — Sol high → Sol light e2e
Full eval path from multi-source corpus through control plane + sandbox + dashboard
Honest prior-vs-Build-Week split so judges score the Codex/Sol delta, not the whole monorepo

What we learned

Frontier models aren’t just autocomplete. With the right loop — Sol high for architecture questions, Sol light for implementation — stuck systems design can move in one session. Education products need both a trustworthy knowledge layer and a way to evaluate it.