Projects

Selected hands-on work. Code and write-ups where I can share them.

GeoCoRD — Causal-augmented multi-agent scheduling

A geo-distributed multi-agent RL scheduler (MAPPO backbone + structural causal models for counterfactual reasoning). Robust under regional outages and 10× bursts where rule-based and centralized-DQN schedulers collapse. Notably, I report an honest negative result: the causal layer, in its original form, hurt reliability — and I traced why, then partially fixed it with a regret-gated penalty. (Under submission.)

Dummy V2 — LLM-agent driven robotic grasping

A 6-DOF arm driven by an LLM ReAct agent: the model reasons, calls motion/vision tools, and grasps objects, with its thinking process streamed live to a web UI. Vision via NPU or classical CV; execution over a real serial/CAN stack. A concrete look at where LLM-planned manipulation actually stands today.

Isaac Sim SDG on Deadline Cloud

A distributed synthetic-data-generation pipeline: NVIDIA Isaac Sim domain randomization scaled out as cloud render jobs (RGB / depth / semantic / bbox outputs). Infrastructure for training data at scale.

More detail and links coming as repos are published.