Is this a model proof of concept?
No. Production AI is one live pathway, with monitoring and a human in the loop. A demo that cannot be rolled back is still a pilot.
Production AI
Production AI for UK and US mid-market firms.
A workflow operators will use on Monday — with an owner, a rollback, and a register your risk team can read.
In short
Production AI at Velosius is an operating problem. We ship one live pathway in your cloud account, with governance, monitoring, and a human in the loop. We do not sell another chatbot demo.
Who this is for
UK and US mid-market operators who already ran a pilot and now need a workflow that survives audit, latency, and a bad Tuesday.
We will not fine-tune a model because a slide asked for AI, and we will not leave prompts, embeddings, or audit logs in a vendor tenancy you cannot inspect.
The pilot is not the product
A summariser in a sandbox tells you almost nothing about Monday. Production AI has an owner, a use-case register, and a way to turn the feature off. If the brief cannot name the person who reviews a wrong output, it is still a demo.
We start with one workflow that already has volume: claims extraction, document triage, ticket routing. The model is the cheap part. The path operators trust is the work.
What stays in your account
Retrieval, orchestration, and logs run in your AWS, Azure, or GCP account, behind your identity provider. The base model can be an API. The prompts, embeddings, and the decision trail do not leave your boundary.
Human-in-the-loop is a rule, not a slogan. Payments, clinical notes, and legal commitments wait on a named reviewer. Amaya Perera leads production AI; Kavindi Jayasuriya reviews the controls on regulated work.
What the first 90 days actually cut
Weeks 1–2 lock the workflow, the data that may enter a prompt, and the rollback. Weeks 3–6 put the thinnest path in your account with monitoring from the first request. Weeks 7–12 sit with the operators who have to live with a wrong answer, then leave a runbook so the next workflow does not need us in the room.
What the first 90 days leave you
- One production workflow with an owner and a use-case register
- Retrieval and audit in your cloud account
- Human-in-the-loop rules for actions that cannot be undone
- Monitoring, rollback, and a composite claims-cockpit snapshot on /work
Work snapshot
Claims cockpit
A UK insurer needed claims handlers out of copy-paste hell. We stood up a governed AI extraction layer on their existing policy stack — human-in-the-loop, auditable, live in 90 days.
First production workflow in one quarter
Longer field note: AI Enterprise: What Comes Next.
FAQ
Before you brief us
Where do prompts and logs live?
In your cloud account. We will use an external model API when that is the right cut. We will not leave the audit trail in a vendor tenancy.
Who leads production AI?
Amaya Perera. Regulated controls are reviewed by Kavindi Jayasuriya. A 15-minute call is enough to see if the workflow is a fit.
Next step
Fifteen minutes is enough.
Send the workflow, the systems it must touch, and the constraint you cannot break. We will say build, buy, or stop.