amazon agi · real-time multimodal agents
technical lead, research tooling · san francisco
leading ten senior engineers across research data, training, tooling, and evaluation. the team shortened the path from a data request to an analyzed trained model from roughly two weeks to three days.
increased a streaming vision-language model’s win rate from roughly 10% to 40% in under a week, and cut qwen 3-vl-8b training from 31 hours to 8.5 hours on a 32-node b200 cluster.
designed a compact computer-use action language and a client-server runtime that reaches under 300ms frame-to-action latency in-region.
