Assistants and agents
Assistants that know your field, your data and your tone of voice: they answer customers, collect requests, prepare drafts and hand over to a person when needed.
Services / AI Integration
A well-integrated AI system is not one more tool to learn: it works inside the processes you already have, takes repetitive work off people's hands and delivers results someone can check. We design it for companies that want to use AI on their own data and in their own systems, not just in a demo.

Experience before the hype
In 2017 we built expert systems for food and spare-parts companies: sales forecasting and risk analysis, the AI of the time. Today we work with language models from Anthropic, OpenAI and Google, and with open-source models installed on the client's own servers when data must not leave the company.
The difference from those who started with ChatGPT is a single one: we know what happens when a system like this has to work with real data, real processes and real people. And we know how to make it work, not just in the demo.

Areas
Assistants that know your field, your data and your tone of voice: they answer customers, collect requests, prepare drafts and hand over to a person when needed.
We connect models to CRM, ERP, management software, calendars and WhatsApp Business through APIs: AI works inside the tools you already use.
Manuals, procedures, contracts and catalogues become a knowledge base you can query in plain language, with answers that cite the document they come from.
Reading and sorting emails, invoices, contracts and forms: the system extracts the data and takes it where it is needed, with a human check where mistakes are costly.
Data from different sources brought together for forecasts and alerts: less time building reports, more time using them to decide.
Open-source models on your servers, cloud services, or both: the most sensitive data stays in-house, the rest uses the most capable models.
What decides quality
The quality of an AI system does not depend on the model alone. The software architecture around it matters just as much: what data it receives and how that data is prepared, which processes it enters, how it integrates with the systems you already use and how results are checked before they reach a customer. That is the part we work on most.
That is also why we build AI into the SaaS products and embedded systems we design: a model is only as good as the system it sits in.
Computer Vision
Cameras and images produce more data than any other sensor, yet they stay silent until someone interprets them. We bring computer vision into the processes you already run: recognising objects and defects, reading documents, counting and measuring, flagging what doesn't add up. When it makes sense, processing happens right on the device, thanks to our embedded background: faster responses, less data on the network, and images stay where they were captured.

Defects, scratches and anomalies spotted on the line, with thresholds you set and an archive of rejected images.
Text, tables and data extracted from photos, scans and forms, ready for your business systems.
People, vehicles and goods counted from cameras, without facial recognition and with privacy in mind.
Lightweight models on cameras and embedded boards, to decide in real time even without a connection.
Real case · Pavé
Pavé receives dozens of enquiries every day by email and WhatsApp, often out of hours and often repetitive. We integrated an AI agent that reads every message, retrieves the details from the database of available properties, replies in a personal way and, if the customer is interested, offers and books a viewing in the agent's calendar.
We estimate that about 70% of enquiries are handled without human intervention, with an estimated response time under 60 seconds, seven days a week. These are our own estimates, not certified measurements.
AI and privacy
There are two different things. Technical security is how the system protects data: who can access it, what reaches the model, what is stored and for how long. We design that. Compliance with the rules, such as the GDPR, professional secrecy or the rules of your industry, also depends on how you use the system and on your contracts: it has to be checked with your own adviser. We don't give legal advice, but we give your adviser all the technical information they need.
With providers' business services, under their terms, the data sent is not used to train their models and is kept only for the limited time needed to prevent abuse. Before we start, we show you the terms of the service chosen. With an open-source model on your own servers, the data goes to no provider at all.
On the technical side, a suitable system can be designed: as little data as possible, personal data removed before sending wherever possible, models on your own servers if needed. Whether, and under which conditions, it is lawful is a matter for the rules: it is the first question to ask your adviser, before starting.
Both are possible. In the cloud you get the most capable models and no hardware to manage, but the data reaches the provider, often in the United States. On your own servers, with an open-source model, the data stays with you, but you need hardware and maintenance. Often the answer is a hybrid.
No system is one hundred percent secure, and no software makes you GDPR compliant on its own. For our own website and apps, the providers and the countries they are in are listed in the privacy policy.
Guides
FAQ
The one that suits the task: models from Anthropic, OpenAI and Google through their APIs, or open-source models on your servers. Often several work in the same system, each on the part it does best.
Yes, it is one of the most common jobs. With RAG the system finds the relevant part of your documents and answers citing the source, so the reader can check.
Yes. That is why we design the checks: answers that cite their source, thresholds beyond which a person decides, automatic checks on extracted data, and quality measured over time.
Usually not. AI connects to what you already use through APIs or through the files your systems already produce.
Cloud models are paid by usage, based on the amount of text processed; a model on your own servers costs hardware and maintenance. Before starting, we estimate the costs on your real case.
Yes, with an open-source model installed on your servers. For some tasks it is less capable than the largest cloud models, but for many business processes it is enough.
AI can probably take it over. Tell us about it: we reply within 24 working hours.