Operator

Theo Park

LLM evaluation · Seoul · 9 years

He builds the holdout the vendor will not like.

FocusThe tests a slide deck cannot survive.

• Holdout design
• Label leakage detection
• RAG and extractor eval
• Happy-path refusal

Closed work2026: Proved the leakDiligence extractor. The answer was in the input. The score was a story. The holdout was not.

Theo designs evaluations for RAG, extractors, and agent fleets. He is the person who puts the real files next to the demo files and writes down the drop. Leakage, vendor-held data, and happy-path scripts are his usual quarry.