San Francisco Bay Area
Research engineer + angel investor with experience across racing, aerospace, automotive, cameras for autonomous vehicles, ML hw/sw co-design (custom silicon), and multimodal foundation models. Gradatim Ferociter (step by step, ferociously)
TwelveLabs is a multimodal foundation model company building SOTA video understanding (VLMs) and video embed, search, and retrieval systems. - Led research and engineering to develop VLM fine-tuning (LoRA/PEFT, HPO, dataset creation and evaluation, LLM as a judge) as a service. - Owned VLM evals (academic benchmark investigations, distributed ray inference). - Contributed to multimodal embed, index, search, and retrieval system.
Investing, scouting, advising, building.
The Machine Learning Hardware Architecture team was a HW/SW Co-Design team responsible for setting the architecture for the Neural Network Engine (NNE) of the Rivian Autonomy Processor (RAP1). Responsibilities included ML acceleration, performance analysis and characterization, compiler development, quantization, and served as the single source of truth for chip direction. As an architect on the team, my primary responsibility were the following: - I developed a performance simulator of the NNE which represented the functional units of the chip - Using the simulator, I mapped and optimized ViT (Vision Transformers) and Wayformer (Prediction Transformers) onto the chip, achieving 90% compute efficiency, hiding nearly all memory movements behind compute. Invented/re-discovered MatMul and GEMMs as a Convolution. - Used the results of the simulation to drive compiler and chip architecture requirements. See RAP1 press release here: https://rivian.com/newsroom/article/rivian-unveils-custom-silicon-next-gen-autonomy-platform-deep-ai-integration https://www.youtube.com/watch?v=nHfKyO9Afj0
During my first year on the MLHW team, I was responsible for bringing up a new compute platform on the vehicle, the NVIDIA Jetson Orin. I modeled, mapped, and characterized multiple deep learning models and architectures onto the chip, including an end to end low latency object detection pipeline. I conducted multiple studies on accelerating ML compute and profiled and improved latency by implementing concurrent programming mechanisms (callbacks, threads, and asynchronous operations). The NVIDIA Jetson Orin and the models that I mapped onto it were shipped with the first major R1S and R1T refresh: https://stories.rivian.com/meet-the-new-r1
I was an early employee of Rivian's dedicated camera team for perception and self driving applications. My primary responsibilities were to bring up the imaging lab, develop in-lab and vehicle experiments, and to develop software for evaluating image quality (color uniformity, relative illumination, SNR, flare) of camera modules. The cameras evaluated and chosen during my time on this team were part of the first major R1S and R1T refresh: https://stories.rivian.com/meet-the-new-r1 In addition to my responsibilities as a camera engineer, I visited the factory in Normal, Illinois several times to support end of quarter production pushes, specifically the end of line (EOL) sensor calibration station.
Primary structure stress analysis and maintenance procedure development
Lots of carbon fiber wet layups, learning about the car, how to use tools (machine shop, hand tools, welding, etc). Competed at Lincoln