Case studies
Agents that work for us every day
None of them is a connector to the cloud. Each sits on our server next to the data it needs: mail, GitLab, the information system. The model is one step in the process, and code makes the decisions.
Shared pattern
Five agents, the same rules
The agents were built one at a time, but they keep the same principles. Those matter more than whichever model they use today.
Data stays with us
The agent runs on our server and reaches into our systems. Credentials come from systemd, not from a config file.
Models only through the gateway
Every call goes through LiteLLM and returns structured output against a JSON schema. The model has no tools; it only proposes or summarises.
Code and people decide
Code decides what happens, by rules. Where an action goes outside, a person confirms it. The mail agent cannot even send.
Input is untrusted
The content of an e-mail or issue enters the prompt fenced off as data, not instructions. Prompt injection has nothing to take over.
Failure is not a quiet day
When a source or the model fails, the agent says so loudly. It never sends an empty digest that looks like “nothing happening”.
No blind retries
An uncertain action is never retried automatically. State and idempotency live in SQLite, so nothing is created or sent twice.
No-obligation consultation
Let’s find the first job for AI in your company.
An hour’s conversation about where your data lives, which processes slow you down and what of it may leave your network. You leave with a proposal for a first pilot, not a tool demo.
Built and is accountable for the AI layer at Nux
Radek Růžička
Managing Director and CTO, Nux s.r.o.