What this is

The unfiltered output of a working lab: everything learned while making self-hosted models do real production work, published because the numbers were too useful to keep private.

The through-line is one genuinely hard workload — a judgment-heavy data-enrichment pipeline that runs for hours unattended, drives a real browser, and gets scored against human-adjudicated ground truth with error weights ratified before any results were seen. Every model faces the same sealed inputs. Every claim on this site traces to an instrumented run.

The lab: a consumer tower with an RTX 4090 and RTX 3090 (and two CMP 170HX cards on the bench awaiting their 64 GB unlock), a MacBook Air as the harness machine, and rented datacenter GPUs when the experiment demands iron the desk can't hold.

The workload's business domain is deliberately anonymized throughout — the methods, numbers, and failures are all real.

Work with me

I consult on exactly this: making open-weights models do dependable agentic work, eval design that survives contact with variance, and the unglamorous systems layer underneath (serving, quantization strategy, thermal/power reality). If your team is weighing self-hosted against API spend, the ledger on the front page is my portfolio. Contact details coming with the launch.