Research
Evidence from building AI in production.
Original studies from our own systems — built on real hardware, run against real workloads, and reported with the full methodology and the limitations. The field runs on demos and benchmarks that can be gamed. We would rather measure what actually happens and show the work.
01 · Field study
An autonomous agent workforce on a single machine
Twenty specialized agents ran 24/7 on one local Mac for three weeks and co-authored a real product's plan and full technical spec. We recorded every run and reported what held up, what failed, and why it failed.
02 · Benchmark study
Does protecting a quantized model's key layers improve agent accuracy?
Two 8-bit quantizations of a 122B mixture-of-experts model, identical except for which layers stayed at full precision, run head-to-head on an agentic tool-calling benchmark. The textbook mixed-precision recipe lost on every metric, and it cost speed.
Why we publish
The work we do for clients is confidential, so we run the same rigor on our own systems and publish the results — including the ones that didn't go the way the literature predicted. It is the clearest way to show how we work: measure first, report honestly, and let the evidence decide. If you have an AI system or decision that needs that kind of scrutiny, that is where a conversation starts.