Data & AI practice
Data and artificial intelligence: from raw data to a model in production
Data engineering and data science form a single chain. A model is only worth as much as the flows that feed it, and a warehouse with no business use is a cost with no return.
Most data projects fail in the same place: between the moment the prototype works on a laptop and the moment it has to run every morning unattended. That transition usually costs more than the prototype itself, and it is almost always underestimated.
We handle the subject in the reverse of the usual order. We start by establishing whether the data exists and is usable, we build the simplest solution that meets the need, and we only add complexity if it beats that reference on a metric agreed in advance.
The symptoms we see most often
- Two dashboards supposed to say the same thing show different figures
- A flow estate where nobody knows which part is still consumed
- One or more AI proofs of concept abandoned before production
- A nightly load window that spills over into opening hours
- A model in production that nobody can confirm still works
- One single person able to change the flows, and they are needed everywhere
How we approach it
Delete before migrating
In a legacy estate, a significant share of flows has no consumer left. Identifying and switching them off is the cheapest gain available, and it mechanically reduces the cost of everything that follows.
One definition per metric
The most common cause of diverging figures is not a bug but two different business definitions of the same metric. We have financial control arbitrate them and we write them into the semantic layer.
The baseline before the model
A moving average, a simple business rule, last month's value. Any model must beat that reference on the agreed metric, otherwise it does not ship — and the reference regularly wins.
Parallel run before cutover
Old and new pipelines run in parallel over the same window, with automated reconciliation. Any gap blocks cutover. We never cut over on the strength of a unit test.
Services in this practice
Frequently asked questions
Where to start when everything looks urgent?
With inventory and lineage. Until you know which report depends on which flow, every prioritisation decision is blind and every deletion is frightening. Two to three weeks are usually enough to get out of that situation.
Do you need a cloud warehouse to do AI?
No. Many useful use cases run on volumes that fit in a properly indexed relational database. The question is not the technology but data freshness, traceability and quality. A cloud warehouse on unreliable data only produces faster errors.
How long before a first usable result?
Framing produces a usable result in two weeks: you know whether the use case is feasible, with which data and at what price. Full production then takes six to twelve weeks depending on complexity and upstream data quality.
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.