4 LLM agents run 108 LoRA experiments over a long weekend
· Hitesh Jain
Four headless Claude Code agents (Architect, GPU worker, Analyzer, Auto-suggester) coordinated through a Postgres queue to run 108 LoRA fine-tuning experiments on gpt-oss-20b. HotpotQA accuracy moved from EM 0.13 to EM 0.72, a 59-percentage-point gain, in ~3 days of LLM time and a few hours of human oversight.