We try a lot, keep a little, and only sell what we use.
The lab is the part of Big Kitty Labs that builds its own products. Work moves inward as it proves itself: from experiment, to a tool our own team relies on, to beta, to something you can buy.
How something becomes a product
Work starts as a numbered experiment: one question, a fixed amount of time, and a written note at the end whether it worked or not. We have run 14 so far. The ones that work usually turn into a script somebody here starts using on real client projects.
If that script is still in use three months later, it becomes an internal tool with a name and an owner. It moves to beta only when someone outside the company asks to use it, and it goes on sale after a full quarter of running for paying users without one of us stepping in by hand.
Products that stop earning their keep are retired in public, with a note on what replaced them.
Recent experiments
Including the ones we shelved.
Reading handwritten site survey forms
Vision model against 300 scanned forms from a real job. 91% of fields correct; checkboxes are the weak point.
Voice notes to pull request descriptions
Worked, but nobody on the team wanted to talk to their laptop. Shelved after two weeks.
Learning a cron schedule from check-in times alone
Seven days of data was enough for 48 of 50 jobs. Became the baseline model in Catnap.
Column type inference without reading every row
Sampling 2,000 rows matched a full scan on all but one test file. Shipped in Tabby v1.6.