Syllabus

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Syllabus

Inspiration

Studying tools often create more material instead of helping a learner return, recall, and improve. Syllabus began with a personal problem: long lectures, generic explanations, and pressure-heavy study plans are difficult to sustain, especially for learners with ADHD.

What it does

Syllabus turns PDFs, PowerPoint decks, Word documents, or pasted notes into a structured tutor. It extracts source text, breaks material into teachable concepts, prepares short morning and evening sessions, accepts typed or spoken teach-back answers, returns source-grounded corrections, exposes supporting citations, and tracks mastery.

The public judge demo provides the complete learning loop with safe sample material and no account.

How we built it

The web MVP uses React and Vite. PDF, PPTX, and DOCX extraction runs on-device with pdf.js, JSZip, and Mammoth. A versioned tutoring contract constrains every turn to an allowed outcome, response, mastery change, and known citation. Malformed or unsupported output is rejected and replaced with deterministic cited review.

Neon Auth and Neon Postgres provide identity and learning data. Vercel hosts the app and serverless routes. Web Speech supports voice with typed fallback. WebLLM provides an optional privacy-preserving local model path.

Codex with GPT-5.6 was the primary Build Week collaborator. It helped turn the original experiment into a deployed MVP, reason through the grounding contract, implement the responsive lesson and judge-demo flows, build the Neon/Vercel architecture, write seventeen automated tests, and verify production.

Challenges

The hardest problem was refusing unsupported text. Citation identifiers have to survive parsing, concept creation, grading, synchronization, and fallback. We made citations part of the response contract and fail closed when a response cites an unknown source.

Accomplishments

• Deployed judge-ready tutoring loop with no account required
• PDF, PPTX, DOCX, and pasted-text extraction
• Strict source-citation validation
• Typed and spoken teach-back
• Morning/evening sessions and mastery tracking
• Seventeen passing tests

What we learned

Grounding is a product behavior, not just a prompt. The parser, database, response contract, interface, and fallback all have to agree.

What's next

Complete immutable local-model distribution, physical-device web-to-iOS handoff, and subscription verification.

Built With

• swiftui