Data science and AI delivered with the MLOps that keeps it alive
Forecasting, scoring, classification, NLP and applied LLM work on your own business data. We deliver the model, the retraining pipeline, the drift monitoring, and documentation your IT department can take over without us.
- Model
- Fixed price or embedded
- Duration
- 6 to 12 weeks
- Expected outcomes
- A model in production with a tracked business metric, not an orphan notebook
- Expected outcomes
- An internal team able to retrain and diagnose without calling for help
Who it is for
- IT departments with data but no in-house data science team
- Business units stuck at proof-of-concept stage for 18 months
- SaaS vendors who want an AI feature inside their product
- Existing data teams overloaded with run work
Who it is not for
- Fundamental research or academic publication
- Projects with no available data and no way to collect any
- "Build us a ChatGPT" with no identified use case
Deliverables
- Use-case framing with a quantified success metric agreed before the first commit
- Exploratory analysis and data quality report (completeness, bias, leakage)
- Trained, versioned model with a documented comparison baseline
- Inference pipeline (batch or real time) deployed in your environment
- Retraining pipeline and drift monitoring (data drift, concept drift)
- Model card: assumptions, limits, validity scope, fallback plan
- Knowledge transfer: two sessions with your teams, commented code
How it runs
01
Framing and feasibility
Two weeks. We audit your actual data, not its description. If the use case is not feasible with the available data, we say so at this stage and you pay for the framing only.
02
Baseline and iterations
We build the dumbest solution that works first. Any more complex model must beat that baseline on your metric, otherwise it does not ship.
03
Industrialisation
Packaging, model CI/CD, prediction regression tests, monitoring, drift alerting. This is the step most POCs never clear.
04
Handover and run
Operations documentation, incident runbook, training for your teams. After that, either you take over or we stay on the model under a maintenance contract.
Stack
- Python
- scikit-learn
- PyTorch
- XGBoost
- MLflow
- Databricks
- SageMaker
- Vertex AI
- Azure ML
- LangChain
- pgvector
How the assignment is structured
Three ways to engage this service. The figure is set at the end of framing, once the scope is written: it is the only way to commit firmly without pricing in a risk you would end up paying for.
Data framing
fixed price, 2 weeks
You have an idea and do not know whether your data supports it
- Audit of the actual data sources
- Quality and volume report
- Feasibility note with target metric
- Firm quote for the full project
Model project
fixed price, 6 to 12 weeks
One framed use case, taken all the way to production
- Everything in the framing package
- Trained and industrialised model
- Retraining pipeline and monitoring
- Model card and knowledge transfer
- 30 days of corrective warranty
Embedded data scientist
time and materials, 3-month minimum
Your data team exists but is short-handed
- Senior profile, 5+ years
- Embedded in your rituals and tooling
- Fluent French and English
- Replacement within 10 business days on mismatch
Frequently asked questions
Our data is a mess. Is that a blocker?
No, that is the norm. The two-week framing exists precisely to measure and price that mess. In half of the cases the real first mission is not a model but data engineering work, and we say so rather than sell a model that will fail.
Can the data stay in Europe?
Yes, and that is our default configuration. Compute and storage stay in your cloud provider's EU region. Our teams connect through a bastion with no local copy, with MFA and access logging. The DPA and standard contractual clauses are signed before the first access.
Do you use proprietary LLMs on our data?
Only if you ask for it, and through endpoints where training on your data is contractually excluded (Azure OpenAI, Bedrock, Vertex). Otherwise we work with open models hosted in your own tenant. The choice is documented in the model card.
Other services
Twenty minutes is enough to know whether we are useful
No sales deck. You describe the need, we say whether it is in scope, at what price and on what timeline. If it is not for us, we say so during the call.