Exploring Computer Science, Machine Learning, and Robotics @Carnegie Mellon

Built PyTorch's first WebGPU backend in ExecuTorch for browser LLM inference with no cloud dependency. Outperformed llama.cpp, WebLLM, and ONNX Runtime on quantized Llama and Qwen across 128 to 8192 token contexts.

First engineering hire. Shipped AI artifact-generation for ServiceRocket and Valiantys, driving up to 80% efficiency gains. Backed by Sequoia's $20M Series A (announcement).

Built FastFax at Stanford Health Care to triage urgent fax referrals with OCR and a CNN. Processed 567,000+ faxes, cutting triage time from 33 hours to 63 minutes at 89% PPV. Published in NEJM Catalyst and as an AMIA poster.