Projects

An Agent Harness system for synthesizing egocentric training data from executable states, semantic anchors, and a temporal bridge layer.

Restoring decision states, executing alternative branches, and verifying continuations from real-world egocentric videos.

LLM preference parsing, multilevel joint auctions, and mixed-integer optimization for charging-station siting and resource allocation.

A video-driven symbolic simulator and belief-state planning benchmark built from HD-EPIC egocentric cooking videos.

A legal QA and agent system integrating Graph-RAG, reinforcement learning, harness engineering, and multi-agent collaboration.