Services

I take LLM features that work in a demo and make them survive production. Agents, RAG, anything that talks to a model and then has to keep running once real users arrive.

What I am asked to look at usually breaks in the same places. A tool call fails and nobody handles it. Retrieval returns something plausible but wrong. The bill grows and nobody notices until the month closes. And there is no way to tell whether last week’s prompt change made things better or worse, because nothing measures it.

Start with an audit

600 EUR, five working days, no call needed.

You send the code, 20 to 40 real questions or scenarios from production, and last month’s API bill. You get back a written report:

• what happens on a tool error, a timeout, an empty model answer, and where the user is left with silence instead of an escalation
• what the model actually receives in context for each of your questions, measured on your data rather than on a synthetic set
• tokens per request, what drives the bill, what can be cut without losing answer quality
• prompt injection, data leaking into logs and third party telemetry, limits that are missing
• the fixes in priority order, each with an hour estimate

If the audit finds nothing worth fixing, I write that. It has not happened yet, but the report is what you are paying for either way.

Then the work

Fixes and features run at 60 to 70 EUR per hour. I keep two clients at a time rather than five, so the work moves instead of queueing. The person who wrote the report does the work, which means nothing has to be explained twice.

How I work

In writing, asynchronously. Specs, reviews and status updates go in text, and I do not hold standing calls. I am in Montenegro, on CET. My written English is fluent; if your process needs daily video, I am the wrong hire and it is cheaper for both of us to know that now.

Contract and invoices go through my own Montenegrin company, in EUR.

What I do not take

Model training and ML research, Python as a project’s main backend, Java, .NET, Kubernetes setup. I say so in the first reply rather than three weeks in.

How to reach me

Email vlad@vitovtsev.site, or use the terminal on the front page: it answers questions about my work and passes a message on to me.