I am interested in the practical stack behind RL for LLMs: rollout generation, reward signals, policy optimization, evaluation, and debugging. I have been exploring open-source RL-for-LLM tooling including rLLM, with an eye toward making post-training more understandable and reproducible.
A judge-driven multi-agent retrieval system for financial document QA. The system decomposes answering into retrieval, reasoning, and judging agents, then uses feedback-driven retries to recover from weak answers. I see this as a bridge between retrieval systems and RL-style improvement loops: feedback, credit assignment, and iterative correction.
My contribution: Contributed to archaeological fieldwork using ground-penetrating radar, LiDAR, 3D scanning, infrared imaging, spatial data analysis, and Gaussian splatting to document and digitally reconstruct the site.
Broader project methods: GPR, magnetometry and electrical resistivity; drone LiDAR, multispectral and thermal imaging; photogrammetry and point clouds; GIS, georeferenced mapping, elevation models and vegetation indices; VR and digital archives.
project / history / methods / Duke overview