Featured work / Meta
Muse Code.
Core developer of Muse Code and contributor to Muse Spark, working across the coding agent and the co-training loop between model and harness.
Muse Code is Meta's terminal coding agent for complex software engineering work across large repositories.
Read the launch- Harness
- Agent runtime
Planning, tool use, persistent agents, validation, and recovery for complex engineering work.
- Co-train
- Model + harness
Develop Muse Spark and Muse Code together so model behavior and agent runtime reinforce each other.
- Long run
- Repository scale
Sustain context, compaction, observability, and reliable execution across extended engineering tasks.
Current work
Agent platforms, from interface to runtime.
Build the system around the model.
Planning, tool routing, context, recovery, evaluation, and debugging surfaces for agents that have to finish real work.
Turn agent behavior into a training signal.
Use harness trajectories, evaluation, and production feedback to improve the model, tools, and product together.
Ship research as reliable infrastructure.
Distributed systems, applied ML, traffic engineering, and product design, with a bias toward legible behavior and dependable operation.
Selected systems
Recent work in agents, ML systems, and infrastructure.
Muse Code
Core development of Meta's terminal coding agent, with contributions to Muse Spark through model-harness co-training.
Agent harnesses and eval loops
Harnesses that let agents use tools, retain context, recover from failures, and improve through measured feedback.
ML systems and network intelligence
Built prediction, optimization, automation, and applied ML systems for infrastructure at production scale.
Research
Selected public research and engineering work.
Background
Recent roles and training.
Meta TBD Lab
AI Research Scientist working on AI agent products and systems.
Netflix
Senior Research Engineer on ML systems and production infrastructure.
UIUC + Tsinghua
Ph.D. and M.S. at UIUC, with undergraduate training at Tsinghua.