One product, one price, one verdict. Anything a machine can do, a machine already does — that part is not what you are paying for.
In the trade, an assay certifies what a piece is actually made of. I offer exactly that for AI systems: fixed scope, fixed price, two weeks — where it breaks under load, what it truly costs per task, whether the architecture survives contact with production. Delivered as a signed written verdict, including the honest one: sometimes the verdict is that what you have is fine.
The judgement is independent, and that is the point. I sell you no models, no cloud, no seats and no licences — there is no vendor whose product improves if the verdict goes one way rather than the other. A dashboard can total your bill; it cannot tell you the architecture is wrong. Every number in this house runs on hardware I own, and carries its date.
The core engagements of this practice are kept under lock — shared only by invitation. If you have been given the access word, you can enter the Private Commissions room →
Every act of AI runs on metered words — what a model reads and writes. Most bills run two to three times what the work needs to cost, and hardware barely touches it: the fix is which model answers which question, what gets cached, what runs in batch, and whether documents arrive as text or as pictures of text. Tell it last month's bill and it returns what the work ought to cost — including "keep what you have" when that is the honest answer.
Reliability under real load, not demo conditions. True cost per task across your actual mix. Failure modes named before they're expensive. Architecture judged by someone who runs his own inference, not someone reselling someone else's. One price, quoted in full before the work starts and unchanged after it — it does not renew. Two weeks, the verdict in your hands at the end, and yours to keep. Taken by measure — judged, as everything here is, by whether the work is the right work.
Each one measures what intelligence truly costs across one real production workload. Anonymized, dated, and banked, they become a benchmark of the true cost of AI across the systems this bench has opened — a standard set here rather than quoted from a vendor. Clients read their own verdict; the record is what the next verdict is measured against.
Land certain, expand into a standing cadence, climb to the sovereign build. The verdict finds the money; the cadence is where it keeps compounding, month after month. The whole staircase is visible from the first rung — that is what makes the first step easy and the last one inevitable.
Fixed fee · proof
Two weeks, fixed scope, one price agreed before the work starts. Where it breaks under load, what it truly costs per task, whether the architecture survives production — described in full above.
Standing · monthly cadence · six months minimum
The verdict, kept true. Setup, then a monthly cadence: which model answers which question, what gets cached, what runs in batch, what moves onto owned hardware — retuned as your traffic, your prices and the models themselves change. A standing relationship, not a project that ends. Setup is waived for anyone arriving from an Assay.
Sovereign build · hardware at cost
Your own inference, on hardware you own, behind your own door. For the business whose thinking has become too central to rent — models, routing and the machinery around them built in-house, handed over whole, with no vendor holding the switch.
Twenty years at one bench. Every line read before it ships, every dependency accounted for, every system signed by the person who built it.
The work arrives finished, and it arrives from one desk — one system at a time, carried from first line to production by the person who designed it.
Most of what gets billed as "AI work" elsewhere is the same request re-read at the most expensive rate available. Finding what doesn't need to be re-read is most of the actual skill — and it's the part that doesn't show up on an invoice.
Every rung above is carried through the same four layers — the intelligence that decides, the systems it runs on, the machinery that keeps running unattended, and the surface your clients touch. One standard through all four, because one person carries all four.
LLM-based systems wired into your data and your tools — sorting, routing, drafting, deciding — with the fault-handling any production service needs, and your judgement about what matters encoded in them.
Off-the-shelf software fits nobody exactly. I build the layer that's yours alone — your data flows, your automations, the connections between the tools you already use — owned outright, not rented. This is most of the actual work.
The part people touch — built by hand, wired to real APIs and real data, fast and correct. The surface where the reputation of everything behind it is settled.
The machinery beneath, the scheduled jobs, and the connections between the tools you already use — built to fail in plain sight, recover without fuss, and keep producing with no one standing over them.
Commissions here are built to order and handed over whole — code, keys, and title. Once delivered, the work belongs to its owner, and so does the telling of it. What can be verified without naming anyone sits one section down, measured and dated.
One company this year receives the full Assay at no charge — the same two weeks, the same signed verdict, nothing abridged. The fee is the permission: the result is published anonymized, as the first entry in the record.
It goes to the most demanding system offered, not the first to write. If your AI spend is already a board matter and you would rather know than guess, make the case in two paragraphs.
This is where the proof lives — one room, one pair of hands, three machines bought rather than rented. A run of 326,531 items, measured and dated; rented cost against owned, plotted; planned against delivered, on the record. No client’s name is needed to verify any of it. The ring behind this page is the cut itself: type, lit, turning — computed as you read, not recorded.
One machine can do the work of a hundred ordinary men.
No machine can do the work of one extraordinary man. Elbert Hubbard, 1899
The mesh does the volume. Deciding which hundred are worth doing is still the job.