Weiqi Fan

Worked on distributed agentic RL training/inference consistency using LLM training stacks. Debugged deterministic MoE routing and reduce-order mismatches across rollout vs training, implementing ordered deterministic MoE forward/backward compatible with different tensor-parallel sizes. Built and integrated parts of an agent training/rollout runtime pipeline, aligning trajectory-to-sample data contracts and adapting Harbor API calls for multi-machine validation and issue diagnosis.

Built a JAX-based multi-GPU differentiable physics solver for complex flow simulations. Re-encoded 3D multi-block grids into continuous vectors and designed multi-layer index mappings to support grid partitioning and vectorized computation. Implemented JIT-compatible irregular multi-to-multi boundary exchanges using mpi4jax and added regression tests to verify numerical consistency across parallel rewrites.

Developed a personal tuning agent that closes the loop between an LLM-generated search space and Optuna optimization. Implemented layered trial search (including TPE/Hyperband) with experiment memory, plus utilities for correctness checks, failure retry, and checkpointed resumption. Reported reductions versus pure Optuna: about 35% fewer trials and about 30% less tuning time, with failed/invalid configurations dropping from ~12% to ~5%.

Contributed to a team project for navigation optimization using reinforcement learning with TD3 and Transformer-based temporal representations. Implemented training and evaluation pipelines and tested policies in complex dynamic obstacle simulation scenarios. Achieved a reported navigation success rate of 86.35% in the project setup.