AI experience
AI features that hold up in production.
A demo that works once and a feature that survives a model update are different pieces of software. The difference is validation, evaluation, and deciding in advance what happens when the model is wrong.
What I build
| AI chatbot | A support or enquiry assistant grounded in your own content, falling back to a contact form when it does not know. Available as a €200–500 add-on. |
|---|---|
| Document extraction | Invoices, contracts and forms turned into validated records, with a human queue for anything that fails validation. |
| Retrieval over your data | Internal assistants grounded in your documents, with citations back to the source so answers can be checked. |
| Workflow automation | Drafting, classification and triage inside the tools you already use, where the model proposes and a person approves. |
| Evaluation harnesses | Golden-set tests that turn a provider's model update from an outage into a reviewable diff. |
How I keep it honest
- Schema first. Output is validated before it touches a database. A failure retries, then escalates to a person.
- Golden sets. Thirty real examples with known-correct answers tell you more than any amount of eyeballing.
- Defined failure behaviour. Before launch we decide what happens with the cases it gets wrong: flag, queue, or refuse to answer.
- Cost and latency budgets. Measured per feature, because a clever prompt that triples response time is not shippable.
- Data boundaries. Which of your data may leave your infrastructure is your decision, written into the design rather than discovered later.
Where it sits in the packages
Basic AI integrations are included in GOLD. A chatbot can be added to any package. Larger work — extraction pipelines, retrieval over internal data, assistants wired to live systems — is PLATINUM, quoted after a scoping call.
On AI-assisted development: I use AI tooling daily for scaffolding, refactoring and review, and it genuinely speeds delivery — part of why these prices work. It does not replace architecture decisions or tests, and I do not ship code I have not read and understood.
When AI is the wrong answer
Plenty of problems that arrive labelled "we need AI" are really a missing report, a badly shaped database, or a form nobody automated. If that is what I find while scoping, I will tell you before you pay for a model.