What we do
The work starts with a simple question: which task is costly today, and does a language model handle it better than conventional automation? Many needs presented as AI projects are solved by a simple rule that is far more reliable and much cheaper.
-
1
Qualifying the use case
Assessment of the expected gain and an honest comparison with a non-AI solution, which is often simpler.
-
2
Preparing the source data
Collecting, cleaning and splitting the documentation that will underpin the answers.
-
3
Prototype and error rate measurement
Test on a sample of real questions, counting the correct and incorrect answers.
-
4
Integration and safeguards
Connection to your tools, a restricted answer scope and human approval on sensitive points.
-
5
Monitoring and continuous improvement
Analysis of real conversations to correct the instructions and complete the knowledge base.
Tech stack & tools
The technologies and platforms we use for this type of assignment.
- OpenAI API
- Claude API
- Python
- LangChain
- Hugging Face
- N8N
Methodology
No deployment without first measuring the error rate on real cases, and systematic human approval wherever a mistake would involve your liability.
What's included
A detailed list of what you receive at the end of the assignment.
- Use case qualification workshop
- Structuring of the knowledge base
- Working prototype tested on real cases
- Documented measurement of the correct answer rate
- Integration into your existing tools
- Safeguards and restricted answer scope
- Conversation logging for improvement
- User training and documentation
Indicative timeline: 4 to 10 weeks depending on the use case and the state of the source data
Training & certifications
The core skills and continuous learning behind this area of expertise.
- Proficiency with the OpenAI and Anthropic APIs and their billing models
- Retrieval-augmented generation (RAG) and document chunking techniques
- LangChain and Hugging Face ecosystems for model orchestration
- Continuous monitoring of the rapid evolution of models and their capabilities
- European regulatory framework for AI systems and the GDPR
Experience & expertise
The kinds of assignments this expertise covers day to day.
- Conversational assistants fed by internal knowledge bases
- Automatic data extraction from unstructured documents
- Integration of language models into existing automation chains
- Setup of quality measurement and conversation logging
Why choose us
An honest opinion on relevance
If conventional automation is enough, we steer you towards it rather than a more expensive AI project.
Error rate measured and shared
You know exactly what the tool gets right and what it gets wrong before putting it in your customers' hands.
Confidentiality decided with you
Together we define which data can go through an external service.
Running costs anticipated
The cost per request is estimated at the design stage, to avoid an unpredictable bill once in use.
Our packages
Indicative prices, adjusted after we discuss your project.
One-off assignment
from €990
First use case
- Needs qualification
- Measured prototype
- Simple integration
- Documentation provided
Project package
from €2,900
Business assistant
- Structured knowledge base
- Integration with your tools
- Safeguards and approval
- Conversation dashboard
- Fine-tuning for 1 month
Ongoing support
from €890 / month
Continuous improvement
- Analysis of real conversations
- Correction of answers
- Knowledge base enrichment
- New use cases
All our prices are indicative. Contact us for a free personalised quote.
Related services
The services this expertise is most often involved in.
Artificial intelligence
Use artificial intelligence to become more efficient and competitive.
Learn moreWeb application development
Custom web tools tailored to the specific needs of your business.
Learn moreExpertise FAQ
Yes, it is the main risk with language models. We limit it by restricting answers to your documentation and measuring the error rate before deployment.
It depends on the architecture chosen. For sensitive data, the model can be hosted in-house, with slightly lower performance but no data leaving your systems.
For an assistant handling a few hundred conversations a month, allow usually less than €30 in API costs. Large volumes are costed case by case.
Rarely. In most cases, retrieval-augmented generation on your documentation gives better results at a far lower cost than dedicated training.
With a set of real questions whose correct answers are known, run before deployment and then again after every change.