Anyone can build you a demo that works.
Two afternoons and an API is enough to show something that answers well in front of a committee.
Production is another matter, and the AI is not what holds it up.
It is access permissions, deciding which version of each document is the right one, and data that cannot leave your network.
An AI system does not break: it keeps answering, only worse.
Which is why trying it for a while proves nothing. You measure it with a case suite that protects every change that comes after.
I take generative AI from demo to production.
With auditing and security included. And if your case does not need AI, I say so before charging you for it.
Three things, and I do them.
No middle team, no reselling. I work on fixed-price projects or monthly dedication, always with working pieces from the first week.
An assistant over your documents
It answers about your manuals, protocols, catalogues or internal policy, always citing the source. What stretches these projects is rarely the model.
AI that never leaves your building
Open models on your server or your GPU. No data travels to a third-party API, which is the only way to work with clinical records or legal files.
An audit of the AI you already have
Case suite, adversarial testing and a failure report with a fix plan. The suite stays yours and protects every future change.
This website is a working demo.
The assistant below is not a form in disguise. An open model answers it, running on my own GPU, with the same guardrails I build for a client: it invents no prices, admits what it does not know, and if it goes down it says so instead of breaking.
Answered by an open model hosted on my own GPU. No data travels to third-party APIs.
What could be built in your case?
Five questions and my own model tells you what it would build, where it would start and what it would not do. If your problem is not solved by AI, it says that too.
This is not theory. It is in production.
Where being wrong is expensive: health data and legal files, with human review and GDPR compliance.
A medical SaaS, built end to end
Co-founder and CTO. Reasoning engine with LLM and RAG self-hosted on a private GPU, a suite of nearly 300 clinical cases, 84 real users across 36 institutions.
A law firm doing 70% less manual work
Drafts of civil claims with overnight processing. The lawyer always reviews and validates: the AI prepares, it does not sign.
Three steps.
Fixed-price projects or monthly dedication, always with reviewable deliveries.
An honest diagnosis
I look at your case and tell you what I would do, how long it would take and what the risks are. If you do not need AI to solve it, I tell you that too.
Building in deliveries
Working pieces from the first week, not promises. You see and validate every step on your own case.
Production and guarantees
Deployment, a test suite that protects every change, security and documentation so your team can maintain it.
Got an AI project on your hands?
Tell me what you want to build, or what isn't working. I'll reply with an honest assessment: what I would do, how long it would take, and whether you actually need AI to solve it.