Keep the steady load. Rent the peak.
Most enterprise AI spend goes on renting capacity for work that runs at a flat, predictable load all year. That is the part worth owning. The rest is worth renting well, on terms written down in advance.
The steady load
The everyday workload: retrieval over your contracts, policies, incident history and drawings. It runs in colocation, a secure room, bare metal or an isolated cloud tenancy configured without training egress.
The peak
Evaluations, fine-tuning, a quarter-end surge. Rented deliberately, with the scope of data that crosses the boundary defined before the first call is made.
The split
Someone has to decide which workload sits on which side, document it, and enforce it in the architecture. That decision, and the engineering that holds it in place, is what we sell.
What we put inside the boundary
A production stack, not a proof of concept. You hold the keys, the logs and the runbooks at handover.
Question answering across contracts, personnel files, incident reports, drawings and operating procedures, with answers that cite the source document.
Ingestion, chunking, embeddings and refresh jobs, kept current as your document set changes.
Task automation with explicit permissions, human approval on consequential steps, and an audit trail of what was called and why.
Open-weight models sized to the hardware, with a gateway that routes each request according to the split you agreed.
We check the licence attached to every model before it reaches production and default to Apache 2.0 and MIT terms, which carry an explicit patent grant and no usage restrictions. Where a vendor community licence is the better technical fit, you see its conditions in writing first.
Identity, network path, retention, logging and monitoring, documented so your auditors and your customers can see how the system behaves.
How an engagement runs
Four stages. Each one ends with something you own, and you can stop after any of them.
Assess
A half-day working session on one real workflow: the data it touches, who may see it, and the controls it needs. You leave with a written architecture and a fixed-scope build proposal.
Design
Models, retrieval, network path, identity, logging and the boundary itself, specified and reviewed before anything is racked.
Build
The stack goes into your environment. Handover includes runbooks, monitoring, a rollback path and the keys.
Operate
Patching, model updates, retrieval maintenance and light on-call under a monthly retainer, or a clean handover to your team and we step back.
Where we fit
Engagements run remotely, in English or Portuguese, wherever your infrastructure sits.
A good fit
- Teams with at least one person or a managed provider running IT
- Sensitive material already in the business: contracts, client and patient records, claims, incidents, drawings
- Legal, healthcare, financial services, real estate, and data center and colocation operations
- A regulator, a customer or a contract that asks where the processing happens
Not a fit
- Training a foundation model from scratch
- A chatbot on a fixed low budget with no data work behind it
- Eighty-page tenders and twelve-month procurement cycles
If the honest answer is that a managed service already covers your case, we will say so on the first call.
Tell us what you want to run
Describe the workflow and where it would live. You get a reply from the person who would run the engagement, not a sequence.
Prefer email? Write to info@keepference.com.