A model is easy. Keeping it right is the job.
Most machine learning projects do not fail at the modelling step. They fail because the table the model reads was built by hand once, or because nobody noticed six months later that the world moved and the predictions did not. We build the pipeline and the monitoring alongside the model, because a stale model is worse than no model. It is wrong with a confident number attached.
This is for you if
- You have years of transaction, usage or operational history nobody has modelled
- Forecasting is done in a spreadsheet by someone applying judgement to last year's figure
- You want to know which customers are about to leave while you can still do something
- Stock, staffing or capacity decisions are made on gut feel and cost real money when wrong
It isn't, if
- Your data lives in an inbox and three spreadsheets. That is a data engineering job first, and we will scope it as one rather than pretending a model can start there.
- You want a dashboard. Dashboards are cheaper and often the correct answer; we will build one and say so.
- The decision would not change whatever the model said. A prediction nobody acts on is an expensive number.
What actually happens.
- 01
Check the data can carry it
Volume, history, label quality, and whether the thing you want to predict is even recorded. This step ends some projects, which is considerably cheaper than the alternative.
- 02
Pipeline before model
The transformations that produce the training table, in code and on a schedule. A model trained on a table someone assembled once cannot be retrained, which means it has a shelf life measured in months.
- 03
Model, then a baseline to beat
Always compared against the simple version, whether that is last year's number or the current rule of thumb. A model that does not beat the spreadsheet is a finding worth having early.
- 04
Ship it where the decision happens
Into the tool the decision is actually made in, not into a notebook. Then drift monitoring and a retraining schedule with a named owner.
What you get
- Demand, churn and lifetime-value models trained on your history
- The data pipeline underneath, because a model on a stale table is a wrong model
- Dashboards aimed at a decision rather than at a screen full of charts
- Monitoring for drift, with a retraining schedule someone actually owns
Built with
- Python
- dbt
- BigQuery
- scikit-learn
- Airflow
The outcome
Decisions made on what the numbers say rather than on what last quarter felt like.
Forecasting, scoring and segmentation on your own data, with the pipeline that keeps it accurate after launch.
Get a free consultationScope and a fixed price before anything is committed. No obligation to proceed.
No dark periods. No surprise invoices.
A structured engagement from the first call to launch, so you always know what is happening and what it costs.
Week 1 · Discovery
Scope & fixed price
Then, every week after
A working demo.
The ones people actually ask.
How much data do we need?
It depends on what you are predicting and how rare it is. Two years of daily history is usually workable for forecasting; rare-event prediction needs enough examples of the event, not just of the data. We check before quoting the build.
Do we need a data warehouse first?
Not always, but often it is the cheaper path, because one place the pipeline can read beats five integrations that each break separately. Where you already have one, we build on it.
Will we be able to explain the predictions?
That is a requirement we design for, not a property we hope for. Where a decision affects a person, as with credit, pricing or eligibility, we favour models whose reasoning can be shown, even at some cost in accuracy.
Who maintains it afterwards?
Either your team, with the pipeline and runbook handed over, or us on a support arrangement. What does not work is nobody, which is how a model quietly becomes wrong.
Often bought alongside
DevOps Services
Pipelines, environments and observability, so shipping on a Friday stops being a brave decision.
Cybersecurity Services
Finding what an attacker would find first, fixing it, and leaving you with the evidence your customers ask for.
IT Infrastructure
Cloud that is sized for what you run, backed up in a way that has actually been restored, and cheaper next quarter.
All services · Working on your own career instead? Career services for engineers
What is the most expensive thing your team still does by hand?
Tell us, and we'll tell you honestly whether software can fix it, and roughly what it would cost. No pitch deck.