Research that survives contact with production

Applied LLM research for contact-center software: hierarchical agentic retrieval, schema-validated model outputs, and the evaluation harnesses that decide whether any of it worked before a client ever sees it.

Most of the job is deciding what to measure. A pre-registered evaluation that disproves your own team's leading hypothesis is worth more than a fix shipped on a hunch, and a result that only holds in a notebook is not a result.

The rest is production forensics — the defect that will not reproduce in validation, the output shape that fails silently on one model family, the retrieval that truncates before anyone notices. I am most useful on ambiguous problems where the hard part was never writing the code.