Product
Built for students who want notes sorted, readable, and easy to revisit out loud.
- Categorizing notesPaste a note and NoteLMs assigns it to a subject so your library stays organized.
- Formatting notesRaw pasted text is turned into clean, structured HTML that’s easier to skim and study from.
- Reading aloudOn Voice, paste text and hear it read back, useful for review without staring at a screen.
Technical details
Under the hood, NoteLMs is also a live experiment in note classification.
- Research goalCompare zero-shot BERT, fine-tuned BERT, and a full reasoning LLM (GPT-OSS 20B) on academic subject classification.
- Eval dataWe’ve already run 2,000 offline evals. Every time someone uses the product, that classification is added to the research set.
- How categorization worksThree models vote in parallel: zero-shot BERT, fine-tuned BERT, and GPT-OSS 20B. An orchestrator (also GPT-OSS 20B) reviews those votes and chooses the final subject.
- How image notes workUploaded or dragged note photos are sent to OpenAI vision, which extracts the raw text; that text then goes through the same classify → format → save pipeline as pasted notes. This is a temporary feature that will be removed soon.
- How reading aloud worksText is split into smaller chunks, then Orpheus 3B (orpheus-3b-0.1-ft) synthesizes speech chunk by chunk so longer passages stay clear and stable.