Who we work with
Who will check your AI evaluation?
An AI evaluation is only as sound as the judge that graded it. When a validator, a council or a customer's reviewer asks how sure you are, an unmeasured judge has no answer. Find your situation below.
For model-risk and validation teams.
A GenAI model reaches validation with an LLM judge's scores in the file. If the validator asks how sure those scores are, a weak answer means a finding, a delay or a model that does not ship.
We check the judge behind one evaluation you already run, on your data or inside your environment, and hand your second line a findings note it can read and challenge.
For clinical AI governance councils.
Your council approves AI tools on the evidence vendors bring. If that evidence was graded by a judge no one measured, the risk stays with the health system after go-live.
We measure the judge behind a vendor's evaluation, or your own, and report each result with its error bar and its date, so the council can see what holds.
For AI vendors selling into health systems.
Health systems that run AI councils ask vendors for evidence before go-live. Evidence the council can check moves the review along. Evidence it cannot check can stall the deal.
We run an independent challenge of your evaluation and write a findings note your buyer's council can read. The note says who commissioned it.
For eval builders.
You run evals and LLM judges, and a score moved. Was it the model, the judge, the item set or noise? The free Jan 7 lesson on Maven starts there.
Two ways to work with us.
- Independent challenge
- One evaluation, one decision, your data. A findings note with a one-page summary for your committee, and a debrief. You choose the turnaround and how small a difference between two versions of your model the test must reliably detect.
- Measurement program
- Your models and scales, measured again when you change a model or a judge. Hosted by us or run inside your environment, both early access by request.
Each fee is set on a 20-minute scoping call. A finer setting needs a larger calibrated question bank, more calibration runs and more checking, so it costs more and takes longer.
Judge credentialing for levels 7 to 13 is in progress. Each credential is published with its level, standard error and conditions of use.
For consulting and Lean Six Sigma partners.
You bring the client and the improvement method, and AIM supplies the measurement inside your engagement. Readings sit on one ruler with error bars, set against your client's spec limits. Where enough repeated readings are in statistical control, results can be shown as process capability (CpK). A single reading gets a precision statement only.
Tell us which number you must defend.
Tell us which evaluation you need checked, and who will review it.
You will get a short note from Matt when a measurement slot opens. Nothing else, no drip sequence.
We store what you send with the few providers our Privacy Policy names, and use it to reply to you. We do not sell it. Email matt@xlnc.co and we delete it. Read our Privacy Policy.