A few engagements, described as plainly as confidentiality allows.

A regulated lender

We ran an AI readiness assessment, then built what it pointed to: an LLM document-extraction pipeline that pulls structured fields from loan paperwork, sends low-confidence cases to a person, and keeps an audit trail for every extraction. It cut manual data entry sharply and left a record a compliance team could stand behind.

An insurer

We built an agentic workflow to triage incoming claims documents: classify, extract, and route, with guardrails that hand anything uncertain to a human. Every decision was logged and explainable, which is the part a regulated insurer actually needs before it trusts an agent with real work.

A healthcare analytics company

We took an ML model that worked in a notebook and made it production-grade: a reproducible pipeline, evaluation against held-out data, drift monitoring, and a quarantine path for inputs that didn't pass validation. It shipped with the controls a regulated setting needs, and the team could explain every prediction.

A data team carrying an ML product

The model worked; the data under it did not. We built the pipeline feeding it end to end, ingestion, validation, reconciliation, and a quarantine path for bad inputs, so the model stopped being retrained on dirty data and the team stopped firefighting every release.