AI models inside your products and your processes
You do not need to train a model from scratch to get real results. You need to work out which part of the job is worth handing to an existing model, how to connect it to your company data, and how to check that its answers hold up.
Applied artificial intelligence means embedding language, vision or forecasting models into software and processes that already exist, connected to a company's own data. It does not involve training a model of your own: it uses models available through APIs and builds around them the context, the controls and the interface that make them useful in daily work.
The problem we solve
Many companies arrive asking to "add AI" without a defined problem in front of them. The useful starting point is the opposite one: find the tasks where time is spent on unstructured material, such as reading invoices and contracts, sorting incoming requests, hunting for one fact inside years of documents, or drafting the first version of a quote or a reply. Those are the jobs where a model earns its place, not the operations already covered by a precise rule.
The second recurring confusion is about training. Building your own model calls for large datasets, specialist skills, hardware and maintenance that never ends. In almost every business case the right route is different: take a model already available through an API and spend the effort on context, meaning which documents and which data the model may read, how the request reaches it, and how the answer is checked. That is where a usable result is won or lost.
Then there is the known issue: a language model produces plausible text, and plausible is not always true. It is kept in check by tying answers to verifiable sources, showing the document behind every claim, narrowing the range of questions accepted, and leaving confirmation to a person wherever the decision carries financial or legal weight. And when a task is deterministic we say so: a database query or a written rule costs less, answers faster and gets it wrong less often than any model.
How we work
An AI project is judged on measured results, not on the demo that worked first time. That is why testing on real data comes before any integration work.
Choosing the use cases
We review the candidate tasks and keep only those with enough volume, with unstructured text or images involved, and with a clear way to tell whether an answer is correct. The rest stay out, even when they would be the most eye-catching to show.
Prototype on your data
We build a trial on your real documents and score it against a set of known cases. We compare models from OpenAI, Anthropic and Google on answer quality, latency and cost per call, before a single line of integration is written.
Connecting company data
We index the documents the model is allowed to consult and build the semantic search that picks the relevant passages for each question. Permissions stay those of your own systems: nobody sees anything they would not already see.
Controls and verification
Every answer cites the source it came from, out-of-scope questions get an explicit refusal rather than an invention, and requests are logged. Where it matters, the model proposes and a person approves before the action becomes final.
Production and cost control
We track spend and response times, pick the right model for each individual operation instead of the largest one for everything, and switch when better versions arrive at the same cost or less.
Technologies
We are not tied to a single model provider. The choice follows the task, the constraints on the data and the cost per call once the system is running.
Models
- OpenAI
- Anthropic (Claude)
- Google (Gemini)
Semantic search
- Embeddings
- Vector databases
- RAG
- PostgreSQL with pgvector
Integration
- Python
- FastAPI
- Node.js
- REST APIs
- Webhooks
- MCP
Interfaces and infrastructure
- React
- Next.js
- Netlify
- Docker
- Serverless functions
Delivered projects
Applications where the AI model is a product feature, not a demo kept to one side.
Allestio · Assistant inside a business system
Within the management system for audio-video companies and exhibition contractors, an AI assistant supports the quoting stage: it reads the client brief, suggests the price list entries that fit the setup, and cuts the time spent putting together a first draft.
CookedThumb · Batch image generation
A web app that turns raw screenshots into polished thumbnails using Gemini. It produces several variants in one pass and lets you refine them by describing the change in words, with no graphics editor involved. Live at cookedthumb.netlify.app.
CLACOROO · Open source desktop application
A desktop tool for managing Claude Code agents and MCP servers from one interface instead of editing configuration files by hand. Open source, available at clacoroo.app.
Frequently asked questions
How much does it cost to add AI to business software?
There are two separate items: building the integration, and model usage, which is billed per call. A first focused feature, such as an assistant over internal documents, starts in the low thousands of euros. The recurring cost is estimated during the prototype, against real volumes.
How long does it take to get a working AI assistant?
A prototype you can judge against your own documents takes a few weeks. A version integrated into company systems, with permissions, request logging and checks on the answers, usually lands in the four to eight week range, like any other software project of similar scope.
Should we train our own model or use an existing one?
In the vast majority of cases, use an existing one through an API. Training demands large datasets, hardware and constant upkeep, and rarely beats a strong general model connected to your own documents. Training is only worth considering for highly unusual data no public model has ever seen.
Where does company data go when it is sent to an AI model?
On paid API plans from OpenAI, Anthropic and Google, the content you send is not used to train the models and retention is time limited. Exposure can be reduced further by filtering data before it leaves your systems, or by running models on infrastructure you control.
How do you stop the AI from giving wrong answers?
By tying answers to verifiable documents and always showing the source, narrowing the questions accepted, and making the system state plainly when it has nothing to go on. Where mistakes are costly, the model drafts the answer and a person confirms it before it is used.
Let us find out whether AI actually fits your case
Thirty minutes to see which activities suit an AI model, which are better left as they are, and what embedding one into your current systems would involve.