The situation
Filevine is a large, established legal-tech platform — thousands of law firms run their cases on it. In early 2024 it had the classic big-company AI problem: real customer demand and nothing shipped. It needed someone who could do the product thinking and the engineering at the same time.
What I did
I embedded as the AI lead and built the customer-facing agent myself: writing the code, setting the technical direction, and owning the product end to end. Instead of pitching a roadmap, I got a working agent in front of people and let the demo make the argument.
Most of the real work sat below the waterline. The flagship use case, AI Medical Chronology, does work firms staff whole teams of paralegals for — reading medical records that run past 500,000 pages and pulling out the data points a case turns on. Getting AI to do that credibly meant improving every stage of the pipeline: better OCR on the scanned records, better extraction of the data points, and a data layer that houses the results so the chat agent can query them.
The eval framework made it sellable: we proved the system's efficacy against the work paralegals were already doing instead of asking firms to take it on faith. Then I taught the firms how to fold it into their practice. Alongside it, I prototyped other AI features across the product and landed each with its owning team.
What shipped
The agent is now Filevine's flagship product, driving about half of the company's ARR — momentum that carried into a $400M raise at a roughly $3B valuation in 2025. More durable than the numbers: a legal-tech org that used to talk about AI now ships it. The teams I worked with kept building after I left.
“The agent was the easy part. The job was making half a million pages of scanned paper legible to it.”