This is not a story about technology. It is a story about a CEO of a contracting company who stopped waiting for an IT department he never had, and built the tools he now sells to other CEOs.
I rented the first server so that a lead from a Facebook ad would not wait until morning. One script, one address, zero DevOps knowledge. It worked. Then came a subsidy calculator that does what a salesperson used to do in their head, and a gateway to language models that every request from my agents passes through today.
start · first VPS · lead webhook · subsidy calculatorAt some point I added up what I was paying for other people’s tools that did not understand my business anyway. Building my own came out cheaper. A GPU server in the office does the heavy lifting. A separate machine takes research waves of several dozen agents at once. Everything is tied together by a private, encrypted network that no key ever crosses in plain text.
then · GPU server · agent farm · private networkSystem ORCA now runs leads, quotes, fund settlements and public tenders in my companies. Over a thousand commits since April 2026. Waves of 48 and 55 tasks without a single failure. Only then did I start selling it. The rule is simple: if I don’t use it in my own production, I won’t deploy it for a client.
today · System ORCA · 100+ agents · ORCA clientsAn agent reads tender specifications, triages email and drafts offers overnight. In the morning your team gets a short list, not a pile.
The cost of one analysis is counted in cents, not salaries. Model routing and limits make the AI bill predictable to the last euro.
Models, data and logs on servers you know. No vendor changes your pricing or switches off a feature mid-quarter.
An agent gets only the data it needs for the task, secrets never enter prompts, and every decision to send, pay or sign is made by a human.
Public services, calculators, webhooks, the LiteLLM gateway, agent memory on Postgres. Runs 24/7.
My own machine in the office: embeddings, indexing, local models, heavy jobs not worth paying the cloud for.
A separate server for research and test waves. A router splits work by capacity, a supervisor watches, a watchdog wakes.
Encrypted tunnels between machines, access only from trusted devices, secrets kept away from agents, backups and logs of every deployment.
I look for solutions that speed up work, but not at the expense of client data or sleep. Speed without control is debt someone will eventually repay. So every system I build follows the same rules.
Process first, agent second. That is why I start with an audit: a list of processes, what to automate, what to leave to people and what it costs. The audit price is fixed up front, not “after a call”. If after 20 minutes I conclude I cannot help, I will say so and point you to someone who can.
Twenty minutes about your company: cash flow, subsidies, AI or entering a new market. If after the call I conclude I cannot help, I will say so and point you to someone who can.
dawid@kawinski.org