Engineering consultancy
Your own AI, on your servers or in your cloud, tested before go-live
We build and run AI assistants and agents on your infrastructure, in your own data centre or your cloud, with your documents and the models that suit your data. Before go-live, an acceptance test shows how often they get it right.
A fixed price for each step, agreed in writing after a free 30-minute call.
Our team has built software and production AI for
On premise or cloud
Your servers, your cloud or both
We help you choose in scoping, before anything is bought. The setup is the same in each case: Kubernetes, a LiteLLM gateway in front of the models and infrastructure as code, so moving later does not mean starting again.
On-premise AIOn your servers
When data may not leave your network, or usage is high and steady enough for your own hardware to pay off.
In your cloud
In an account you control on AWS, Azure or the provider you already use, when a quick start or changing load matters more.
Both, behind one gateway
When different data needs different rules: sensitive work on your own models, the rest on hosted models in an EU region.
How it works
Three steps, each with a fixed price
You get a written proposal after the first call, and each step can stand on its own.
- 01
Scoping or acceptance test
One or two use cases with a measurable goal, and a sizing of hardware or cloud capacity. If you already have an AI system, an acceptance test instead, ending in a written report.
- 02
Build
The assistant or agent and the platform underneath, in milestones, with acceptance criteria written into the contract.
- 03
Run
Your team runs the system after handover, or we do under a monthly retainer with regression tests and a quality report.
What we build
Assistants and agents on your own documents
Each one starts with a use case that has a number attached: hours saved, answers found, cases prepared. We build it in your systems and hand it over with training.
Service assistant
Answers from manuals, spare parts lists and past tickets, for technicians and the hotline.
Quotes and tenders
Requirements, earlier offers and the right product data, found for each new quote.
Process agents
One step of a process prepared, such as checking an incoming order, then handed to a person.
A small model for one task
Trained on your data for a frequent task, and cheaper to run than a large model.
Platform and operations
The platform underneath, run like the rest of your infrastructure
One gateway for every model
LiteLLM with a key, a budget and logs for each team, and rules for which application may use which model.
Model serving
Open-weight models on GPU nodes with vLLM, sized from measured load.
Deployments and monitoring
Terraform, Helm and CI/CD in your repositories, with monitoring and alerts for answer times, errors and cost per team.
Run by us or by your team
A monthly retainer with a named engineer, or a handover with runbooks and training.
Testing
Tested before go-live
A demo shows that a system can answer. An acceptance test shows how often it answers correctly: 100 to 200 cases from your own documents, a required pass rate for each criterion agreed in advance, and a confidence interval around every result.
| Criterion | Required | Cases | Measured | Confidence interval | Result |
|---|---|---|---|---|---|
| Answer is correct and complete | 85% | 200 | 94.0% | 89.8% to 96.5% | Met |
| Cites the right document | 90% | 200 | 95.0% | 91.0% to 97.3% | Met |
| Hands over to a person when it should | 90% | 120 | 94.2% | 88.4% to 97.1% | Not decided yet |
The third criterion is not decided yet: the required rate lies inside the interval. At the same pass rate, about 200 cases would settle it.
Clients
What clients say
“They were embedded in our team, which was really effective.”
“We are impressed with their engineers' high-quality work.”
“Their responsiveness, friendliness, and successful project delivery were top-notch!”
- 01
Your environment, your rules
Every data flow is agreed with you before work starts.
- 02
A data processing agreement in every contract
With the documentation your data protection team needs.
- 03
Teams and tasks, never individuals
Works council requirements are handled before anything goes live.
- 04
Everything we build stays yours
Code, test sets and runbooks, written for your team to run.
Agentic coding
From assisted coding to bounded autonomous workflows
Autonomy expands one task type at a time. Your team starts with assisted coding, moves to supervised agent tasks and hands narrow, well-tested work to background agents only when pilot results support it. Each step has approvals, fallbacks and stop conditions agreed in advance, and decisions rest on measured cost per accepted change. Your engineers keep review and release authority throughout.
Works with
LLM cost optimisation
Spend by team and task, then budgets, defaults and routing.
Harness engineering
Context, checks and review rules that make agents reliable.
Coding agent adoption
One shared setup and training on real tickets.
Evaluations on your repositories
Replays of merged changes before you switch tools or models.
Team
The people behind the work
Mickey GrafFounder and CEOFrom the founder
Models get cheaper every quarter. Running them well doesn't
New models arrive every few months. Whether they pay off depends on the work around them: context, checks, routing and measurement. That is the work we do.
Mickey GrafFounder and CEO
Ask AI for a summary of Cloudsail
Pick a topic and open it in your assistant.
Summarise what Cloudsail (cloudsail.com) does, which services it offers and how a first project with them runs.
FAQ
Common questions
What does Cloudsail do?
We build AI assistants and agents that run on your own servers or in your cloud, set up the LLM platform underneath, and test both before go-live so acceptance rests on measured pass rates. After go-live we run the system under a monthly retainer, or hand it to your team. For engineering teams we also roll out coding agents.
Should our AI run on premise or in the cloud?
It depends on your data, your usage and how quickly you want to start. Your own servers suit data that may not leave your network and high, steady usage; a cloud account you control suits a quick start and changing load. We compare both in scoping, and the setup is the same either way.
Do we have to buy hardware?
Not at the start. A first use case can run in a cloud account with GPUs in an EU region, and move in house once usage and results justify the hardware.
Who runs the system after go-live?
Your team, with the runbooks and training we hand over, or we do under a monthly retainer with a named engineer.
Where do we start?
Where AI already costs money or where work still moves by hand. We measure that first, and the first project answers one clear question.
What does a first project look like?
One process, team or model, with the measures agreed before we start. We take a baseline, make changes one at a time, test each against it, and hand over the setup with training.
Are you consultants or engineers?
Engineers. The people who scope your project build it with your team and hand it over.
Will this work with our existing codebase and systems?
Yes. We work in the repositories, tools and processes you already have, without a rewrite or a new platform.
We already pay for Copilot, Cursor or Claude Code. Why has nothing changed?
Licences help individuals. Results come from the setup around them: the context agents see, the checks before merge, budgets and routing, and training on real tickets. That is the part we set up.
Do your agents decide on their own?
Only within limits you set. Each agent starts by preparing work for a person to decide, and takes on more only after you have reviewed its results with us.
Which models and tools do you use?
Whichever suit the task: commercial models such as Claude, GPT or Gemini, open-weight models on your own hardware, and the coding agents your teams already have.
Can everything run on our servers?
Yes, for agents, test sets and open-weight models. Commercial models run only through data flows you have approved.
Who handles data protection and the works council?
We handle your works council's requirements before anything goes live, and prepare the documentation your data protection team needs.
What does it cost?
It depends on the scope and the size of your teams. You get a written proposal after the first call, and the first conversation is free.
What if a change doesn't pay off?
Then we leave it out. You keep the test set and the measurements either way.
Start with one use case
Bring one use case or the AI system you need to accept. 30 minutes with an engineer, no access to your systems needed.
Look for the draft in your email program
Send it from there, and an engineer replies in person with times for a 30-minute call. If no draft has opened, write to us at miki@cloudsail.com.