Research
We publish the parts that are useful.
No launch posts, no benchmark victory laps. Technical writing on how applied AI is evaluated, where they quietly fail, and what we've had to build to find out.
What we're working on
Thread 01
Evals that survive a model swap
Grading the job, not the prompt; frozen case sets; catching a tool that has learned the test instead of the work.
Thread 02
Reading hurried photos
Shelf and site photos taken at an angle, under bad light, by someone in a rush. What it takes to get a reliable answer from them anyway.
Thread 03
AI for markets that run on WhatsApp
Thai-speaking shoppers, home buyers messaging at midnight, voice notes, half-English messages. Tools for the way business is actually done outside the US.
Tell us the job your team hates. We'll tell you if AI can do it.
A 30-minute call with the people who would build it. No technical knowledge needed: just describe the job, how often it happens, and how you check it's done right.