Services
Applied AI & automation
I bring AI in where it genuinely helps: assistants, agents, RAG search, classification and automations that remove repetitive work. No flashy demos: measurable solutions, with cost control and continuous evaluation so they keep working over time.
How I work
- 01
Sort the wheat
We look together at which task you want to improve and whether AI is the right tool. Sometimes the answer is no, and I'll say so before we start.
- 02
A measurable trial
The narrowest case that can be evaluated, on your real data, with a success criterion agreed in writing.
- 03
Production with guardrails
A budget per query, tracing and continuous evaluation. Nothing reaches users until we know what happens when it's wrong.
- 04
Measure and decide
Weeks later we review the numbers: if it helps, we expand; if not, we stop. Expanding blind is how the money goes.
Where AI genuinely helps
AI is good at tasks where language is the problem: searching scattered documentation, classifying ambiguous input, drafting text, extracting data from unstructured prose. It's bad (and expensive) when used for something a query or a rule handles better.
The first conversation usually serves to separate those two. It's common to come out of it with a smaller scope than you walked in with, and that's normally a good sign.
What I work with
Language models from several providers, chosen by task and cost. Semantic search with RAG over vector databases. Agents with tools when the job takes several steps, connected through MCP to the systems you already use. And orchestration of flows that combine model, rules and code, because the model alone almost never solves it.
How I build it
I start with the narrowest case that can be measured: one task, one set of real data, a success criterion agreed before any code is written. With that working and evaluated, expanding is cheap; without it, any expansion is blind.
For search and assistants I work with retrieval-augmented generation (RAG) and one rule I don't negotiate: the system doesn't make things up. If it can't find support in your documents it says so, and when it answers, it links the source.
What I check before starting
Whether the data is in usable shape, who maintains it, what happens when the model gets it wrong, and what it costs per month at real volume. Without those four answers, a pretty demo never reaches production.
What's included
Answers with a source
Retrieval with verifiable citations. If the system can't link where a fact came from, it doesn't assert it.
Cost under control
A budget per query, caching where it makes sense, and the smallest model that does the job.
Continuous evaluation
A case set that runs on every change. Without it you can't tell whether a prompt tweak broke something else.
An exit when it doesn't apply
Sometimes the right answer is a SQL query, a rule, or solid RAG instead of expensive fine-tuning. You'll know in the first conversation, not halfway through the project.
Want to bring in AI without the hype?
Tell me which process you want to improve and we'll assess whether AI is the right tool.
Schedule a call