Applied AI
Applied artificial intelligence for business problems
We integrate language models and machine learning into existing operations: assistants that answer over your own documents, automatic email and ticket classification, demand forecasting and fraud detection. We start from a concrete business metric and only ship if the model beats the baseline.
The usual mistake: starting from the model
Most AI projects fail because they begin by choosing technology instead of a problem. We invert the order: identify a task with measurable cost, establish the current baseline, and only then choose the approach. Sometimes the correct answer turns out to be a rule rather than a model, and that is a valid result too.
Retrieval over your own documents (RAG)
The most common case. We index your manuals, contracts, historical tickets or internal policy into a vector store — we use pgvector on PostgreSQL, to avoid adding another piece of infrastructure — and the model answers citing the source passage. The answer carries its reference, so whoever reads it can verify it.
Cases by sector
Applications we have built or have a defined architecture for.
- Retail: demand forecasting and dynamic repricing
- Health: triage assistance and drug interaction alerting
- Finance: risk scoring and transaction fraud detection
- Manufacturing: predictive maintenance and logistics optimisation
- Services: automatic ticket classification and routing
- Legal: clause extraction and contract comparison
Privacy and control of your data
When data cannot leave your infrastructure, we deploy open models inside your own environment. When it can, we use commercial providers under a no-training agreement on your data. That decision is made during architecture and written down, because it drives both cost and performance.
How we measure that it works
Every deployment carries an evaluation set of real cases with expected answers. We run that evaluation on every model or prompt change, the same way an application runs its tests. Without it, updating the model is a bet rather than a decision.
FAQ
What does running an AI assistant cost per month?
Cost depends on volume and model. An internal assistant handling a few thousand queries a month typically costs between 20 and 150 USD in model calls. The dominant cost is almost never the model: it is the work of keeping the document base it answers from up to date.
Can the AI make answers up?
It can, which is why design matters. We constrain the model to answer only from retrieved documents, require it to cite the source, and explicitly permit it to say it does not know. That does not eliminate the risk, but it reduces it to a measurable level we verify with the evaluation set.