Audio tools I design and ship as an AI-native engineer. I write the spec, set the architecture and verify the result; Claude Code agents write most of the code under a review gate I built, with separate agents checking logic, code quality and mutation testing.
A 12-stage pipeline where nine LLM subagents write lyrics and Suno style prompts for commissioned catalogs. Linguistic quality gates check rhyme, metre, cross-batch diversity, hook uniqueness and content safety, and regenerate whatever fails.
End-to-end voice-over editing: Whisper large-v3 transcribes, Claude marks retakes and dead air through structured tool use, and the tool writes a REAPER script that makes the cuts. Confident edits apply automatically; uncertain ones are flagged for review.
A local manager for a sound library of 800,000+ files. Hybrid keyword and semantic search is benchmarked as a spike so far.
A neural renderer that turns MIDI into an instrument performance. Stage one, MIDI-DDSP in a pinned CUDA container, renders end to end on an 8 GB GPU, starting with cello. A second refinement stage is planned.
A sample-discovery tool built at a hackathon. I built the analysis layer: librosa, Essentia and LAION-CLAP, with a calibrated confidence score on every tempo, key and tag estimate. I also added filename BPM and key parsing as a fallback.
A Chromaprint matcher that finds the same recording across two archives using offset-aligned similarity, plus an EBU R128 loudness normaliser.