I'm an AI product engineer at Google. On nights and weekends I build AI products that get people together in person. This page is the system behind that work: how I run research, how I build, and the things I've shipped.
Not a philosophy. These are the moves behind every project below, and they are the same moves whether the stakes are a party game or a decision I can't take back.
For anything that matters, I don't ask for an answer. I spin up a dozen or more agents that argue opposing positions, then a critic that tries to tear each one apart, then one more that attacks whatever they agreed on. Only the ideas that survive a real fight get through. The rule I hold hardest is to attack the idea I like most, the most. I have killed my own products this way.
I pay a flat rate for these tools, so saving capacity is leaving value on the table. Most people try to get an answer cheaply. I would rather run fifteen agents and get a deeper, better-stress-tested one. More compute buys more rigor.
I have a full-time job, so I can't babysit. Once the destination is clear I hand the AI a long-running objective and a loop that keeps it working and re-planning as it learns. It makes the obvious calls itself and only stops me for the decisions I can't undo.
Every time the thinking changes direction, the system records what we chose, what we rejected, and why, with a date and a version tag. The test is simple. A stranger should be able to read the history and reconstruct the whole story.
I built it a memory. A personal knowledge base that every new conversation inherits: what I'm working on, who's who, what got decided last week. When I switch tools it writes a clean handoff so the next one is up to speed in a few minutes.
A real thing live tonight beats a perfect thing never. Red Knot Club went from nothing to live in one night. Every project gets the same one-command deploy, so there is no gap between built and shipped.
I won't read a wall of raw files. Every conclusion comes back as one clean page with the bottom line first and the single hardest idea turned into something you can click. The plain files are for the machine. The page is for me.
This is roughly what runs when I research something hard. Move the intensity. Each node is an agent with its own point of view. They explore in parallel, argue, and collapse into one answer.
A real question goes in. Something with stakes, where a fast answer would just be a guess. The first job is to frame it into something worth a dozen arguments.
Faithful visualization, no live model calls. The controls and the pipeline below drive the same simulation.
All public, all live, all linkable. The figures below are real test counts and run lengths, not marketing. Click any one to go see it.
Some of my favorite builds are private, so they are not on this page. The pattern is always the same. Someone I care about has a high-stakes decision, so I build them a dashboard for it. Real options pulled live from the source, scored against a rubric I can defend, every number labeled with where it came from and when, and anything I could not verify marked unverified. No model sits in the data path, so it does not make things up. CFI, above, is the public version of the same craft.
Find every real option that exists right now.
Pull each one live from its source.
Drop what doesn't meet the hard constraints.
Rank against a rubric I can defend.
Label every number, flag the unknowns.
4 of candidates · rubric v3
every cell traces to a source and a date · unverified rows never get a rank they didn't earn
redacted mock · fake data, real structure
The point of all this is not the machine. I build AI so the people around me spend less time wrangling software and more time in the same room. Camp Neja, a weekend with twenty friends. Pregame, the games we actually play at parties. The tools are how I get there.