Inspiration
Students often leave a lecture with notes that look complete but quietly miss an important explanation, correction, or connection. Traditional summarizers create new notes, but they rarely show whether the student’s existing notes are complete or where each claim came from. We built LectureWeaver to make note improvement auditable, evidence-linked, and focused on completeness.
What it does
LectureWeaver compares lecture materials, transcripts, and existing Markdown notes. It identifies concepts that are covered, partially covered, missing, or contradictory, then links every finding to trusted page or paragraph evidence.
It calculates a deterministic coverage score, rebuilds the material into clearer enhanced notes, creates a navigable table of contents, generates reviewable Markdown changes, and produces Anki-ready cards. Users can copy or download each output. An optional live workflow supports multiple AI providers, languages, audio transcription, and AI-generated study guides, while the included demo works without an API key.
How we built it
We built LectureWeaver with Next.js App Router, TypeScript, Tailwind CSS, Zod, PDF.js, Vitest, and Testing Library.
PDF, TXT, pasted text, and Markdown parsing happens locally in the browser. The application normalizes each source into structural chunks with application-owned identifiers and trusted locators. Zod schemas validate model responses, while additional semantic checks reject unknown evidence, invalid source combinations, malformed patches, and incomplete outputs.
Live analysis uses server-side provider adapters for OpenAI, DeepSeek, and Kimi. OpenAI uses the Responses API with strict Structured Outputs. Scores, evidence hydration, Markdown generation, Anki exports, and tables of contents are calculated deterministically in application code.
Challenges we ran into
The largest challenge was making AI-generated analysis trustworthy. Models could return well-written content while referencing invalid chunks or producing structurally correct but semantically incomplete results. We solved this by separating generated reasoning from trusted source metadata and failing closed whenever evidence could not be resolved.
We also handled inconsistent PDF extraction, provider-specific JSON behavior, long-running requests, optional input combinations, browser file limits, and Vercel’s audio payload constraints. Another challenge was clearly communicating the difference between local processing, simulated demo analysis, live provider requests, and temporary API credentials.
Accomplishments that we're proud of
We are especially proud that the no-key demo exercises the real validation, extraction, chunking, fingerprinting, evidence-hydration, scoring, and export pipeline.
Every displayed locator and excerpt comes from freshly processed source chunks rather than generated text. LectureWeaver also goes beyond detecting problems: it turns the audit into complete enhanced notes, exact Markdown changes, and source-grounded Anki cards.
The application is responsive, multilingual, deployable on Vercel, and remains fully demonstrable without paid API access.
What we learned
We learned that reliable educational AI requires more than a strong prompt. Strict schemas, semantic validation, deterministic calculations, and clear ownership of evidence are essential.
We also learned that identifying missing content is only the beginning. Students receive much more value when the application transforms findings into clearer notes and practical study materials. Finally, provider-compatible APIs are not truly interchangeable, so each provider needs its own adapter, validation strategy, and error handling.
What's next for LectureWeaver
Next, we want to improve support for longer lectures and recordings, expand document-format compatibility, strengthen multilingual output quality, and add richer Anki export options.
We also plan to evaluate LectureWeaver against larger synthetic and instructor-reviewed datasets, improve accessibility, and explore personalized review plans based on missing concepts. Before enabling publicly funded live analysis at scale, we would add authentication, rate limits, usage quotas, monitoring, and cost controls.
Log in or sign up for Devpost to join the conversation.