Hi, I’m Assaf.

I’ve spent twenty-five years in software engineering, and I’ve followed AI closely as an enthusiast for as long as the field has been worth following. I’ve built with the tools, watched what they could and couldn’t do, and paid attention to where the work has been going.

A year ago, I found myself in a different kind of work: my own family-law case in Oregon.

I knew before I started that chat AI wasn’t going to help me with the substance. It was going to help me draft the occasional letter, and that was about it. The real problem wasn’t writing. It was reach. I had years of evidence in front of me, emails and messages and financial records and court filings and scanned documents, all of it in formats that didn’t talk to each other, and I knew the answers I needed were in there. I just couldn’t get to them.

What made the work heavier than I expected was the cost of going back through it. Reading the same difficult exchanges twice while looking for one specific moment. Pulling tax returns I’d filed years ago and trying to remember what was happening that quarter. Spending evenings inside the worst chapter of my life, looking for the one detail that mattered for a hearing two weeks away.

This is the part of a legal case no one talks about. Your lawyer never lives it. The paralegal isn’t there at midnight. The case is yours, and so is the work of going back into it.

So as I was working through my case, I started building the system I needed.

Not as a separate project. As part of the work itself. Each thing I needed to know, what was actually in the W-2s, what the messages said about a particular week in 2022, where the contradictions were across years of filings, I built the part of the system that could answer it. The pattern was one I knew well from software engineering: agentic AI organized around a reasoning framework, where the language model plans and writes but never retrieves, never calculates, and never validates its own output. Code does that work. The model is kept in the lane it’s good at.

By the end of the case, the system was finding things I’d missed. It was doing in an afternoon what would have taken me a weekend by hand. And every answer it gave me came back the way an engineer wants software to behave: traceable, grounded, and honest about what it didn’t know.

When the case closed, I showed it to my partner Gwendoline, who runs a divorce coaching practice in Portland. Her response, more or less:

Every parent I see needs this and doesn’t know it exists.

Gwendoline Van Doosselaere, Artemis Divorce Coaching

CasePilot is my attempt to package what I built for others who need it. I built it because I wish I’d had it when I needed it.

— Assaf