Useful AI in a product reads real data, returns answers your code can trust and never acts on its own without a check. I add OpenAI-powered features to existing products with guardrails, evaluation and a budget.
Who it's for
- Products that want an assistant over their own orders, documents or catalog
- Teams automating manual work: extraction, classification, drafting
- Founders who tried a demo and need it reliable in production
Problems I solve
Answers you can't trust
Free-form text, invented facts, nothing your code can validate.
Agents that act too freely
An assistant that can write to the database or send messages without confirmation.
Costs that grow with traffic
Uncached calls to the most expensive model for every request.
What you get
- Assistants with tool calling over your real data — read freely, write only after confirmation
- Structured outputs validated by schema, so the rest of the system can rely on them
- Retrieval over your documents or catalog where it actually helps
- Test sets and checks to measure quality before and after changes
- Caching, model routing and spending limits
How it works
- 01
Pick the job
One task where AI saves real time or money, with a measurable before and after.
- 02
Prototype on real data
A small test set from your data decides whether it's good enough.
- 03
Guardrails
Permissions, confirmation steps, validation and fallbacks when the model fails.
- 04
Ship and measure
Quality, latency and cost tracked after launch.
Related work
Read on the blog
Questions
Which models do you use?
Mostly OpenAI models via their API, chosen per task; the architecture lets you switch providers.
Is my data used for training?
API data isn't used for training by default with the major providers; we also keep personal data out of prompts where it isn't needed.
Can AI take actions in my system?
Yes, through explicit tools with permissions — and for anything that changes data or money, after a human confirms.
