Daniel Cadeau

Software Engineer

Artificial intelligence

Use AI on your data, inside your tools.

I start with one specific need and test it on a sample of your data before building anything. We measure what it can actually do and what each run costs.

What I build

Extract data from your documents

We define the information to retrieve, then the system reads each document and always returns the same structure, for example a JSON file ready to import. Expected formats and values are checked before the data is sent to your tool.

Example: turn supplier invoices into rows ready to import into your accounting software.

Query your documents and databases

You ask a question the way you would ask a colleague. In documents, vector search finds the relevant passages and the assistant cites its sources. In a database, it generates a read-only SQL query and turns the result into a clear answer. It only accesses data the user is allowed to see.

Example: ask what notice period applies to a contract, or get quarterly revenue by region.

How much notice do I need to give before ending the master agreement?

Vector Search Tool

Master-agreement.pdf#page=120.91

Terms-and-conditions.pdf#page=80.84

The agreement requires 60 days’ notice.

Master-agreement.pdf#page=12

Can you give me Q3 revenue, broken down by region?

SQL Tool

SELECT region, SUM(revenue) AS sum
FROM sales GROUP BY region

Sure — here is the breakdown by region:

Region	Revenue
North	184 200 €
West	126 750 €
East	98 430 €
South	74 210 €

Agents and automation

An assistant can work inside an existing process: read a request, find information in your tools, call an API and send the file to the right place. The actions it may take are defined in advance, and a failed step does not trigger the next one.

Example: read a customer request, complete it with CRM information and send it to the right team.

Test an idea before building it

Before building the complete solution, I test the idea on a sample of your data. You find out what works, what gets in the way, what each run costs and whether a conventional automation would be simpler. This lets you compare several ideas and choose or rule out the one worth prototyping.

From the first test to production

Audit

I analyse the need using a sample of your data. You receive a summary explaining what is feasible, what data is needed, the estimated cost per run and any simpler solutions worth considering.

Prototype

I build one limited workflow and we test it on cases with known answers. The criteria are agreed beforehand, so the result tells us whether to continue, rethink the idea or stop.

Production

If the test is conclusive, I integrate the solution into your tools with the necessary checks, execution logs and documentation. You receive the code and access.

How your data is handled

Before choosing a model or hosting provider, we decide which data may leave your system, where it is processed and when it is deleted.

Data processing governed by Article 28 of the GDPR

I only process your data to carry out the work. The agreement states which data is involved, why it is processed, which providers are used and when it is deleted.

Choose where processing takes place

Depending on the sensitivity of the documents, we can use a general provider’s API, a provider hosted in Europe or a model installed on infrastructure you control. The choice affects cost, maintenance and sometimes answer quality.

Limit what is sent

We list exactly which documents and fields the model needs. Everything else stays out. Temporary files created during processing are deleted at the end of the run.

Questions & Answers

Is the data used to train a model?

No: this is one of the selection criteria. During the audit and prototype, I use services whose terms exclude using your data for model training. If the project moves to production, the provider and hosting setup are then chosen with you and documented in the agreement covering that stage. The most sensitive documents can remain on infrastructure you control.

What does it cost to run?

There is the development cost, then the cost of using the model. That second amount mostly depends on the volume of text processed. I measure it on a sample of your data before development to provide an estimate based on observed results.

What level of reliability can you expect?

I do not quote a reliability rate before evaluating the solution on representative cases. We prepare examples with known expected results, compare the responses with those results and analyse the errors. You can then decide based on measured results rather than an isolated demonstration.

Do I need a lot of data to start?

No. A few dozen representative documents are often enough for a first test. The aim is not to train a model, but to see how an existing model behaves on your data.

What tools do you use?

I choose tools based on the project’s constraints. This can include Ollama to run a model on controlled infrastructure and FastAPI to connect the solution to your applications. I can also build directly on the technologies already used in your project. The solution can be integrated into the existing application or run separately: the choice depends on confidentiality, volume, cost and maintenance.

Let’s talk about your project.

Book a first call

30 minutes to talk through your project, understand what you need and answer your first questions.

Book a slot