At Beatdapp, I built and led development of a user-level fraud model serving 500M monthly active users. It flagged 2.9B suspicious streams in one year at 99% precision and saved clients $1.7M.
I designed and built Beatdapp’s AI music detection capability from scratch, translating academic research into custom CNN and transformer models. The system scored about 30M tracks for the company’s largest client with a false-positive rate below 0.1%.
For Content ID, I led development of an audio fingerprinting system in Python and C++, testing algorithmic and neural approaches. The neural model identified three-second clips with up to 98% accuracy and was designed to scale to 150M songs.
I also designed ML pipelines that cut GPU training cost 12x and implemented production monitoring for data ingestion, model outputs, and drift. Earlier, as a Software Engineer Intern at Pometry, I built temporal graph models with Raphtory.
