Most of the work is getting language models and machine learning into existing software: retrieval, agents, evaluation, and the data pipelines underneath. I design the system and write the first production version.
Teams rarely struggle with choosing a model. They struggle with data access, permissions, evaluation, latency, and who owns the system after launch. That is the work I take on.
The demo looks fine in a meeting. There is no measure of quality, latency, or cost under real use, so it stays a prototype.
Documents, tickets, and wikis sit in several tools. People want answers with sources, and security needs permissions respected.
Incoming files or records that someone reads, classifies, and re-enters. Extract, validate, and write back, with a person in the loop when confidence is low.
Every request hits the largest model. Routing, caching, and smaller models usually cut cost without changing what the user sees.
Scoring, forecasting, ranking, or image classification. A trained model on your data is often cheaper and more reliable than an LLM.
Build versus buy, model choice, retrieval layout, cloud vendor. A second opinion before you commit budget.
Scope and fee are agreed in writing before work starts. Most clients begin with the audit. If we continue within 30 days, that fee is credited against the next engagement.
What to build, whether it is worth doing, and what it would take.
A design your engineers can implement and own.
One workflow, in your stack, measured against a baseline.
Ongoing architecture once a system is live. A second day a week is €4,500 more per month.
Ranges depend on scope. After a short call you get one fixed price in writing. Model and hosting usage is billed at cost. Fees include VAT.
We pick one workflow, agree how success is measured, and stop if the evidence does not support going further. If a language model is not the right answer, I will say so on the first call.
The problem, the current systems, and whether this is a useful fit.
Data, infrastructure, and process reviewed. Written recommendation.
A design, an evaluation plan, and a costed backlog your team can build from.
One workflow on your data, measured against a baseline, then monitored and handed over.
From the first call through to production, you talk to me. I design the system and write the code. Nothing is passed on to someone else after we start.
I build AI pipelines that stay useful after launch: monitored, straightforward to change, and not tied to a single model version as the stack moves on.
Quality, latency, and cost targets are agreed up front. If a change cannot be measured, it does not go out.
Code, runbooks, and dashboards stay with your team. You should not need me to keep the system running.
Thirty minutes is usually enough to decide on an audit, a build, or that this is not the right next step.