An open, $0-per-user version of Maestro — an AI tutor that runs entirely in your browser. No login, no server, no per-lesson cloud bill. The model runs on the learner's own device, so it's free to scale to the planet and works offline after the first visit.
Live: https://openmaestro.vercel.app · North star: everyone, everywhere, in great quality.
The hackathon's hard rule is $0 COGS — which forces a small, on-device model. Small models are mediocre tutors for lack of discipline, not IQ. So instead of a bigger brain, we wrap a small one in a behavioral teaching engine that handles the exact situations Maestro slips on (don't leak the answer, don't rubber-stamp wrong work, catch the discouraged student, be exact with math). Measured in a reproducible in-browser eval: the engine lifts the model on those scenarios.
pnpm install
pnpm dev # http://localhost:3000No API keys, no .env, no database — there's nothing to configure. On first visit the browser
downloads the model once (~2 GB, from Hugging Face's CDN) and caches it; after that it loads
from disk and works offline.
pnpm test # unit tests (progress, verifier)
pnpm test:e2e # Playwright: offline + persistence
Everything runs in the user's browser — two CDNs seed it once (app + lessons, model weights), then the teaching engine, the on-device model, and browser storage do the rest. No server, no database.
- Inference on-device via wllama (llama.cpp / WASM), CPU — runs on any device, no server.
- App + lessons + quizzes are static files on a free CDN.
- Model weights come from Hugging Face's CDN, then browser-cached (OPFS, kept via
navigator.storage.persist()) — downloaded once, ever. - Progress, streak, chats, profile live in the browser (
localStorage). No backend, no database, no accounts — every new user brings their own compute. - Offline via a service worker; instant-start boots a tiny 1B in seconds then silently upgrades to the 3B.
- Always-on discipline (
lib/tutor/prompt.ts) — explain/answer by default, withhold only when the student should produce the answer; signal shifts; show before you test. - Per-turn skills (
lib/tutor/skills.ts) — cheap deterministic detection injects one focused instruction per moment (distress → empathy, challenge → nudge, submission → don't rubber-stamp, math → be exact, …). - Verifier (
lib/tutor/verify.ts) — a $0 pass over the model's reply that recomputes its arithmetic and corrects wrong numbers, and hard-blocks answer-leaks in challenge mode. - Mastery gate — an objective MCQ quiz the tutor opens (no self-marking, unique per attempt); the lesson only completes when it's passed.
- Remembers the student — name + progress, injected into the prompt from message one.
/eval runs the 20 TutorBench scenarios on-device and scores each with a deterministic
rubric (no LLM judge → $0, offline, transparent). "Prove the engine" runs the same model
with and without the engine and shows the lift, scenario by scenario — measurably better
teaching from the same brain, which is the whole thesis.
app/ Next.js App Router pages (tutor + /eval)
components/ UI: onboarding, lesson chat, sidebar, control center, globe, quiz
lib/llm/ engine.ts (wllama adapter) + registry.ts (model tiers)
lib/tutor/ the teaching engine: prompt, skills, verify, progress, quiz, assessor…
lib/eval/ scenarios, deterministic rubric, run harness
data/syllabus/ the three courses (Python, Business, Frontend) as JSON
docs/ DECISIONS · DEMO · PRESENTATION · SCENARIOS
docs/DECISIONS.md— the full design dossier (what/why/how, business model, tradeoffs).docs/DEMO.md— the 3-minute demo script.docs/PRESENTATION.md— the 10-minute talk outline.docs/SCENARIOS.md— paste-able prompts to show the tutor handling TutorBench live.
It's a 3B running on CPU in a browser — occasional quirk, and the first visit downloads the model once. Cross-device sync isn't built (state is per-device by design — private and $0). The eval rubric is a consistent proxy, not a human grader. The behavior — won't cheat, won't rubber-stamp, catches the math — is the moat, not raw polish.