Approach
From a free workshop to a measured result in production.
We prove value on your real data before you commit, integrate into what you already run, evaluate before and after shipping, and leave you with a system that is documented and owned.
- +Start with a real sample, not a slide deck
- +Fix the success metric before the pilot
- +Keep a human decision on sensitive actions
- +Evaluate before shipping, monitor after
01 — Engagement phases
Decisions made in the right order.
Each step is small enough to decide quickly and solid enough to build on. You can stop after any of them with the deliverables in hand.
Discovery workshop · free, 90 minutes
Bring your product, your process, or a sample of around twenty real requests or documents, anonymized if you prefer. We show what is feasible today, what it would take, and you leave with a first estimate of the gains.
Diagnostic · 5 to 8 days
Architecture and process review, three to five use cases quantified in time, volume and quality, prioritized roadmap, GDPR and AI Act classification. Fixed price.
Pilot · 4 to 6 weeks
A focused workflow on your real data, with a success metric agreed before we start and an evaluation set built from your cases. Runs in your environment, in the cloud, or on our GPU infrastructure in France when the data must not move.
Build and integrate · 8 to 12 weeks
Production integration with your systems, human validation steps where the risk requires, evaluation gates in the delivery pipeline, dashboard, hardening, documentation and training of your internal owner.
Run and improve · subscription
Hosting, monitoring, model and prompt updates validated by evaluation, support, a monthly dashboard and a quarterly review of gains with management. New use cases are added on the same base.
02 — Design principles
Five rules we do not negotiate.
Human in the loop
No offer, answer or action leaves without a person's validation where the stakes require it.
Integration, not replacement
The solution plugs into your ERP, mailbox and tools. Nobody changes how they work to accommodate the AI.
Data under control
Hosting matched to the data, anonymization before any external model, a clear map of what moves where.
Evaluated, not assumed
Every system has an evaluation set and metrics; a model or prompt change ships only when the numbers hold.
Measurement from day one
Response time, volume processed and quality are tracked from the pilot and reviewed every quarter.
03 — Deliverables
Outputs that align technical and business stakeholders.
Diagnostic report and roadmap
Processes, volumes, quantified use cases, priorities, expected gains and the compliance classification of each case.
Pilot decision summary
The metric agreed, the metric measured, what changed, remaining risks and the path to production.
Architecture and integration design
Components, data flows, hosting model, integration points, access control and operating assumptions.
Evaluation set and harness
Real cases, expected outputs, metrics and the automated checks that gate every change to prompts and models.
Governance and compliance checklist
Access patterns, approval points, logs retained, GDPR processing records, AI Act obligations and risk owners.
Handover documentation
What runs where, who owns what, how to update models and prompts, and how to extend to the next use case.
04 — Engagement formats
Four ways to start.
Best when
Fixed-price diagnostic
You want to decide what to build or automate with numbers, and a roadmap you own whatever you do next.
Best when
Fixed-price pilot
The use case is clear and you want proof on your real data before committing to a build.
Best when
Build and subscription
You are ready to put a system in production and want it hosted, evaluated and improved over time.
Best when
Embedded senior engineers
Your team needs senior .NET, Azure or AI capacity for a mission, in your tools and process.
05 — Common questions
We don't have a data or AI team.
That is who we build for. We design, integrate and maintain the system, and train one internal owner on your side.
We have a product team; we need capacity, not a project.
Embedded engineers join your team by the day, in your tools and process, for two to twelve months, with knowledge transfer built in.
Our ERP is old or very specific.
We connect to what exists through its database, files or APIs, build reusable connectors and MCP tools, and quote unusual integrations separately.
Can our data leave the company?
It does not have to. Personal data is anonymized before any external model, and hosting follows the data: public cloud where it fits, France or on-site where it must.
What if the pilot misses its metric?
The metric is agreed before we start. If it is not met, you keep the findings and the evaluation set, and decide with the measured results in hand.
What about the AI Act?
Each use case is classified during the diagnostic. Commercial automation cases are minimal-risk in principle; transparency obligations are handled in the design, and the evaluation evidence doubles as documentation.
Which model will you use?
The one that passes the evaluation on your tasks under your constraints. The architecture keeps the model, the hosting and the framework replaceable.
Bring the product or the process that is slowing you down.
A feature to ship, quote requests to answer, documents nobody has time to read: the first step is a 90-minute workshop on the real thing.