We build it in production. Then we hand it over.
No discovery phase that produces a slide deck, and no phase five.
- Applies to
- Fixed-fee engagements
- Your time
- A handful of interviews, then one engineer part-time
- Exit
- Agreed in writing before we start
Map
We interview the teams already using AI, whether or not that was sanctioned, and read the proxy and DLP logs with your security people. That tells us which workflows are worth automating and which data is involved.
Size
Hardware is sized against real concurrency and context length rather than a vendor's reference architecture. Three or four open-weight candidates are benchmarked on your tasks, at the quantisation you would actually run.
Build
One workflow someone performs daily, shipped properly: retrieval over your own documents, the automation that acts on the output, logging, rate limits and an evaluation set your experts graded. We pick the boring, frequent job over the one that demonstrates well.
Hand over
We pull a GPU node, expire a certificate and corrupt an index on purpose. Your engineers fix all three while we sit behind them. Anything they cannot fix yet becomes a documented gap with an owner and a date.
What you keep
Sizing memo
Two hardware options, arithmetic shown, so your engineers can check it.
Cost model
Three years against your current API spend. A spreadsheet, not a slide.
Golden set
Work your own experts graded. It outlives us and the model.
Runbook and drill
The manual, and a recording of your team recovering from a failure we caused.
Fixed, and what moves it
Fixed fee
Quoted against a written scope. Misjudging the effort is our problem, not a change request.
What moves it
Air-gapped sites, more than one data domain, an identity estate mid-migration, regulator evidence.
Never billed
Scoping calls, your security questionnaire, travel in [REGION], or telling you not to hire us yet.
Terms
[X]% on signature, balance on delivery, [NET] days. You contract hardware directly, at your prices.
Asked on every first call
Open weights vs frontier
Behind on hard reasoning. Level on the summarising, extracting and drafting that is most internal work.
How much hardware
Less than most quotes assume. A department often runs on one multi-GPU box. See the sizing table.
Air-gapped sites
Yes. Models and updates arrive on physical media with checksums. Slower build, longer runbook.
Is private cloud enough
Usually. A single-tenant VPC satisfies most clauses; some programmes need the machines in your building.
Who runs it after
Your platform team, with tools they already use. No such team? Say so early and we will tell you to wait.
Is this just a chatbot
Rarely. Most of the value is automation: intake, extraction, routing and writes into the systems you already run, with an approval gate.
Will you sign our NDA
Yes, before the first real conversation. Security posture and insurance certificate on request.
Send us the workload that cannot leave your network.
Fixed fee, and you keep the memo either way.