research

I helped train and scale CSM’s Implicit Shape Foundation Model, a 3D-native latent diffusion-Transformer model that produces 3D occupancy fields from single image in <1 minute, and its pre- and post-training data infrastructure. My work helped CSM land #1 on the 3D Arena Leaderboard and close initial enterprise contracts.

I helped train and scale CSM’s Neural Radiance Field (NeRF) Foundation Model, a 2D-native NeRF model that reconstructs 3D occupancy fields from single image in <1 minute. This model plugs into a slow refinement process based on Poole et al., 2022 and grew CSM to over 300k+ users. My work was presented at NVIDIA GTC 2024.

I developed adversarial methods for end-to-end text-to-speech models as part of my Master’s thesis. I also lectured on these concepts for Stanford’s CS 236G: Generative Adversarial Networks Winter 2021 offering. Here are some samples from my implementation, which is an adaptation of Donahue et al., 2021; Kim et al., 2020.

I trained and shipped models to the v9.2, v10.1, v10.2, and v10.3 releases of Tesla’s Full Self-Driving (FSD). My work was also featured in Tesla’s 2021 AI Day event.

I trained and shipped lightweight speech vocoders to the v1.0.0 release of NVIDIA’s Neural Modules (NeMo), an open-source speech AI toolkit for developers that has now expanded more broadly to conversational AI and LLMs.