The problem
Voice notes solve capture and fail retrieval: you transcribe once and never find it again. And sending your meeting audio to the cloud trades privacy for convenience.
The product
MDRecord captures system audio and microphone at the same time (the meeting on one side, your voice on the other), transcribes, generates an LLM summary and indexes everything in a local SQLite. None of your notes needs a network to be found.
Benchmark, not adjectives
Instead of claiming “better search”, the repository publishes a reproducible held-out benchmark:
Search blends semantic ranking with SQLite full-text (FTS5) — the hybrid wins because each side fails in different places: semantic misses exact terms, textual misses paraphrase.
Why it matters
Your notes never leave the machine. Ranking is auditable (scores logged locally). And the algorithm can be iterated without API cost — that is how the hybrid was calibrated.
My role
Product author: capture, transcription pipeline, indexing, hybrid search and the benchmark that validates it.