AI systems that don't make things up
Assistants and automations that answer from your business data, know when they are unsure, and hand the decision to a person. Built on the Claude API and measured before a single line changes.
What we do
We build systems where a language model does part of the work and your software does the rest: assistants that answer over chat, WhatsApp or SMS with the business's real information; automations that read free text and turn it into data; integrations that connect the model to your CRM, your database or your billing.
The model is never left on its own. We surround it with rules, verified data and a person who decides when it can't. That is the difference between a demo that impresses and a system you can leave running in production.
How we use it
Four rules we apply to everything we build with AI:
- The model proposes, the system decides. We ask for structured output and make the model choose among options that come from your database, never calculate or invent. If the number isn't in the data, it doesn't go out.
- When it's unsure, it escalates to a person. A "we'll get back to you shortly" is worth more than a confident made-up answer.
- Measure before you touch. Every prompt change runs through a battery of real cases against the real model; if it gets worse, it doesn't ship.
- Cost is capped by design. Small models where they're enough, spend ceilings per user and per day, and never asking the model for what a database query can answer.
An AI system we built
A nutritionist in the US needed her clients to log what they eat by text message. The model doesn't calculate calories: it picks among candidates from an official nutrition database and, when it can't resolve the entry, escalates to her. Each text is acknowledged in 128 milliseconds, and her clients use it every day. Read the full case.
And the one on this page
The chat in the bottom-right corner is Dot, built with the same tools we sell: it runs on the Claude API, answers questions about what we do, and is instructed not to invent facts. Only if you give it your name and email, and agree, does it send us a summary of the conversation. The model makes that call through a tool, and the server validates it before anything is sent. Try it.
Who decides
We use AI throughout our process, and a lot of it. But what gets built, how it is designed and what reaches production is decided by a person who understands your whole system and answers for every line. Who is behind it.
How we work
Discovery
We talk about the problem and the data you have. No jargon, no empty promises.
A trial on your data
Before building, we test with real cases whether the model solves what you need and how accurately. If it doesn't, we tell you.
Build
The whole system around the model: data, rules, escalation to people, costs. Short iterations, visible progress from day one.
Launch
We ship it, measure what it answers and what it costs, and improve it together.
Before you start, worth knowing
- Not everything needs AI. If a rule or a database query solves it, we tell you before charging you for a model.
- Your data trains nothing: the model receives just what each answer needs, and the accounts stay in your name.
- The cost per request is known before going live and limited by design.