Projects

Parallel coding agents that coordinate through a shared context and task queue instead of a main agent. State-of-the-art on Terminal-Bench 4.0 and DeepSWE: up to 17.5 points more accurate than the strongest baseline and 2.49× faster than Claude Code. I also built the open-source DeLM plugin for Codex and Claude Code.

arXiv 2026

Keeping the largest attention weights is already near-optimal for KV-cache eviction; our training-free ContourKV wins 93 of 160 paired comparisons against the state of the art

NeurIPS 2026 Workshop on Long-Context Foundation Models

Consolidating dynamics-guided steering into an unguided diffusion robot policy, reaching up to 91.8% strict success on CALVIN tasks

Stanford CS224R

Test-time learning framework enabling LLMs to accumulate and reuse reliable problem-solving strategies

Stanford CS224N

Multi-agent traffic simulation integrating VQ-GAN camera embeddings via cross-attention

Stanford CS231N

Robust spatio-temporal ozone pollution forecasting via GPR model

Published in JPUR 2023

Gazebo + PX4 drone simulator with RRT-based path planning

Published at AIAA '21 Conference