I’m bringing FreeToken, FlashML’s 15K-star inference engine, to Macs. I got a 122B model running on a 36 GB Mac at 5 tokens a second by streaming it off the SSD with read-ahead, and wrote Metal kernels that make a 35B model generate 1.6x faster than llama.cpp.
Efficient robotics with Coleman Hooper in Kurt Keutzer’s group. I’m finding where robot policies fail: I ran 500 simulated LIBERO trials, then pitted Claude Opus 5.5 against a local vision model at pinpointing the moment each failure starts.
I work directly with NVIDIA, benchmarking their models on Allus’s video inspection workloads. I mapped the Pareto frontier of quality, speed and energy on Blackwell GPUs, and cut latency 62%.
I shipped Truth24, a personalized AI news app that’s made thousands of dollars in revenue, and built PromptLabs, an AI video tool that took production from 4–5 hours a video to 10–15 minutes.
I moved a client onto open models running on an NVIDIA DGX Spark, cutting their AI costs 10x and doubling their speed. I also built a lip-sync video pipeline that cut the cost per video 90% and closed a five-figure contract.
I won their hackathon, then joined. I trained a transformer from scratch to predict which users become power users, which drove $25K+ a month in revenue, and cut inference costs $45–50K a month by moving to Kimi K2 on Groq.