I build AI products that turn foundation models into reliable systems—LLM agents, real-time voice, evaluation, and human feedback.
Building Candor — human feedback infrastructure for evaluating and improving LLM agents.
Human signal collected where AI builders already work—ratings, comparisons, interviews, and structured feedback for agent evals and RL workflows.
↗Agent rollouts, simulator fidelity, judge calibration, generated-state feedback loops, and the conditions under which simulated experience transfers.
—Real-time speech-to-speech translation designed to disappear into live conversation. Co-founded and acquired by Cisco/Webex.
↗A real-time music visualizer that maps audio frequency features into a temporally smoothed walk through a generative latent space.
↗A learned spectral mask for suppressing non-speech noise and improving intelligibility in everyday home and office environments.
—Before meeting platforms exposed usable bot or audio APIs, I built the now-common participant-bot pattern for Zoom, Meet, and Webex using containerized adapters and virtual microphone, speaker, webcam, keyboard, and mouse devices. The approach is covered by US 10,911,718.
—Android OS, kernel and HAL integration, OTA infrastructure, app distribution, and security work for real payment hardware.
—Why spectrograms make audio models feel familiar when you already understand computer vision.
A small MNIST experiment exploring sparse expansion as a learning prior, inspired by the fly olfactory system.