The company
Signal Foundry Labs is a product research lab in Toronto. It’s a small team, lean and mighty - a founder, a few collaborators, and a crew of AI agents - and all of it is pointed at software for work that has to be right.
Signal Foundry Labs is founded and run by me, William White, with a small team around me and more on the way. The bet is that AI can now generate work at almost any scale, but the judgment that says which of it is right still lives in people, and there isn’t much of it to go around. So we build systems around the people who have it: they do their job inside a working app, every accept, fix and rejection becomes ground truth, and a crew of agents turns that into a measurably better version overnight (or tries to - the record is what says whether it worked).
The product is Tenali, an open-source review queue and eval layer. Your agent files the cases it can’t call, you judge them in seconds from the keyboard, and every verdict becomes a label with a pass rate and an honest confidence interval behind it. It runs on your machine as one SQLite file in your project. Tenali plugs into any agent that speaks MCP, tuned for Claude Code and Codex, and a service can call its Python API directly. It never calls a model itself, so it does not care which one wrote the output.
We run our own work through the same loop. The daily updates, the accuracy gates, the plain changelog - that’s how the company operates day to day, not just what it sells.
How we build
Agentic engineering
Coding agents implement, test and iterate around the clock while people direct. That's the only reason a daily improvement cadence is affordable for a team this size.
Measurement first
The people who do the work already know what a right answer looks like. We build software that learns from them, keeps score on itself, and says plainly which work it can handle and which still needs a person.
Daily loops
We ship, measure, and improve every day: a verdict given in the afternoon shows up in the product next morning, and no change ships without beating the last version on the record.
I’ve spent about 25 years building ML systems for media companies, Pandora and SiriusXM among them - mostly recommendation and personalization, and more recently transcripts and talk content. When the tool I need doesn’t exist I tend to build it, which is where the seven issued US patents came from. Every system I’ve built was supposed to learn from its users. The difference now is the loop can run daily instead of quarterly, and it scales with tokens instead of engineering time.
The workflows I care about are the ones where a demo that works 80% of the time is worthless. I design the systems, and the evaluation loops, that go after the last 20%.