Desktop audio manager & AI mastering suite
Product Design & DSP / ML Engineering
Electron audio manager pairing LUFS/true-peak guardrails with CLAP/MERT mastering ranking.
Sole engineer across the Electron/TypeScript desktop workbench and Python/TypeScript DSP & ML evaluation pipeline (audio-core).
Used for managing large AI/original audio libraries, batch stem workflows, and side-by-side mastering candidate auditioning; DSP core published at github.com/abrahamahn/audio-core.
Engineering & product case study
Problem & workflow context
Reviewing dozens of mastering chains or generated takes by ear suffers from loudness bias (louder sounding "better") and inconsistent true-peak clipping before streaming-platform delivery.
Decision & rejected alternative
Enforced deterministic ITU-R BS.1770 LUFS loudness matching and true-peak ceiling guardrails first, using CLAP/MERT perceptual embeddings only as a secondary tie-breaker among compliant candidates. Rejected ranking purely by neural aesthetic scores, which routinely favored clipped, over-compressed masters.
Failure encountered & diagnosis
Batch-analyzing hundreds of long stereo files initially stalled Electron UI responsiveness and spiked memory when decoding PCM buffers concurrently. Diagnosed via worker heap/IPC profiling and moved FFmpeg/DSP measurement into bounded background worker queues with streamed summary metrics.
Verification, role & current status
Verified LUFS/true-peak calculations against synthetic sine/step reference signals with known dBFS/LUFS targets and tested level-matched A/B auditioning across multi-hundred-track playlists.
Architecture & reproducible evidence
- Level-matched A/B mastering auditioning eliminates loudness bias during human review.
- Hard true-peak and integrated LUFS guardrails reject clipped renders before ML scoring.
- Bounded background DSP worker queue keeps Electron playlist browsing responsive on large catalogs.
Technology stack
See all 9 roles, Penn M.A.S. status, U.S. citizenship/relocation logistics, and printable résumé on one page.
Open Recruiter View & Public Résumé →