Language models
- Claude API
- OpenAI API
- open-weight models where required
AI systems
AI for real work. Under engineering control.
Products, internal tools and AI agents for startups, growing companies and creator businesses. We connect models, data and systems. Engineers verify the result and own the delivery.
What we build
-
Products, internal tools and MVPs
Web applications and tools for the team: dashboards, forms, search, working consoles. AI sits inside the process rather than bolted on as a chat box. We start with a prototype that shows the thing working on data agreed in advance, which can be a representative sample or synthetic records; the production version is a separate agreed scope, with tests, deployment, monitoring and maintenance where maintenance is agreed.
-
Agent workflows and request handling
Inbound email, forms and tickets are read, classified and routed to the right person with a draft reply. Anything unclear, sensitive or high value goes to a person. The system keeps logs at the level the task warrants, so the decisions it makes can be reviewed and corrected.
-
Documents, research and company knowledge
Retrieval over documentation, procedures and past answers, plus briefs, reports and first drafts. Where an answer is meant to rest on your documents, it points back to them, the evaluation questions are written before rollout, and judgement on the result stays with the team.
-
Integrations and data pipelines
Systems that were never designed to talk are connected through APIs, webhooks and scheduled jobs. Data is cleaned up and stored where the team actually looks for it. Each step has a defined failure path: a retry, an alert or a handoff to a person.
-
Production delivery and maintenance
Containers, release pipelines, environment configuration, logs, metrics and alerts. Model costs are budgeted and tracked. It is the same delivery work we run for cloud and automation, set up for AI systems.
-
Technical discovery and second opinions
A review of an idea, an architecture or an existing rollout: what can be built, what is missing in the data, where the risks and dependencies are. It ends with a written recommendation and a scope that can be costed. Sometimes the recommendation is a plain automation or a process change rather than AI.
Who controls what
An AI system is only as trustworthy as the controls around it. These are agreed before the first agent runs.
- Scope and success criteria
- Written before the sprint starts: which process, what counts as done, and how the result is checked.
- Data and access
- Agents work with the minimum data and permissions the task needs. Rules for customer data, retention and model providers are agreed with the client.
- Review and tests
- Automated tests and human review check output before release. Evaluation sets are kept, so later changes can be compared against earlier behaviour.
- Deployment and handover
- An engineer approves each production change through the route agreed for that environment. The client receives code, prompts, configuration, runbooks and the agreed access; third-party accounts and licences transfer only where the vendor allows it.
Who this is for
Funded startups, growing companies and creator businesses with a concrete need: a product to build, or a process that costs real hours. It takes a person on the client side who makes decisions, and the budget to do it properly.
We work as a pipeline: design the solution, orchestrate the models, tools and integrations, verify the result against the data we agree, and hand it over documented. A prototype shows the thing works; the production version is a separate agreed scope.
How we scope it
- We start with one product or one process that can be described and measured. The rest follows in later sprints.
- Prototype and production version are agreed separately, so it is clear what is a demonstration and what is a solution to maintain.
- Decisions on money, legal matters and customer accounts are approved by a person. The system prepares and recommends; it does not close them on its own.
- If a plain automation or a process change solves it more cheaply, we say so at the consultation.
Tools in use
Agents
- tool calling
- agent orchestration
- evaluation sets
Retrieval
- Qdrant
- PostgreSQL
- embeddings
Workflow
- n8n
- webhooks
- scheduled jobs
- REST APIs
Runtime
- Docker
- Kubernetes
- GitHub Actions
Observability
- structured logs
- cost tracking
- Grafana
Questions about AI systems
Where does this start if we only have an idea?
With technical discovery: what can be built on the data you already hold, what is missing from it, and where the risks and dependencies are. It ends with a written recommendation and a scope that can be costed. Discovery can also be the whole engagement, and sometimes the recommendation is a plain automation or a process change rather than AI.
Is a prototype the same thing as the production version?
No. A prototype demonstrates the approach on data we agree in advance, which can be a representative sample or synthetic records rather than your live systems. Production readiness is a separate scope: tests, deployment, monitoring, and maintenance where maintenance is agreed. A proof of concept can also stand on its own.
What does an agent decide on its own?
As much as we agree for that process, and no more. Each agent runs inside permission limits set with you, is evaluated against the task it was given, and keeps logs at the level that task warrants. Consequential steps, money, legal matters, client accounts, go to a person for approval.
What has to be in place on our side?
Someone who makes decisions, and access to the data and systems inside the scope. Agents work on the minimum data and permissions the task needs, and that access can be withdrawn. A commitment to production is not a precondition: discovery, or a proof of concept on its own, is a complete piece of work.
Request a consultation
Describe the problem, the outcome that matters and any deadline that is real. Two or three sentences are enough to start.
Copy this text if you need to contact CloudWarrior directly. Answers are kept only on this open page.
Scripts are off, so this form cannot hand the request on. The same three answers work as plain mail: address, company and what needs to change. Write to pat@cloudwarrior.io.
Or write straight to pat@cloudwarrior.io.
Who answers, and when
Status: Patryk Ciszewski is currently contracted on a project.
- Andrzej and Zosia run the initial qualification, 08:00–16:00 Europe/Warsaw. They ask about the company, the project, the problem, the outcome you want, the deadline and the budget, then pass the case to an engineer.
- Technical consultations with Patryk are held after 18:00, Europe/Warsaw.
The time is agreed by email, case by case.
A person reads every request and replies by email with a proposed next step, the questions still open, or a clear no.
What happens next
- A person at CloudWarrior reads the request and checks whether the problem fits the practice.
- The reply proposes a call, asks what is still unclear, or explains why the fit is wrong.
- Andrzej and Zosia run the initial qualification 08:00–16:00, and consultations with Patryk are held after 18:00, Europe/Warsaw. The actual date is agreed individually.
- Scope, success criteria, timeline and cost are agreed in writing before any work starts.
- Reply
- By email, from a person
- Call windows
- Qualification 08:00–16:00 · Patryk after 18:00, Europe/Warsaw
- Dates
- Agreed individually
- Commitment
- None until scope is agreed
Contact
Remote across the EU · Europe/Warsaw time