Demand forecasting
Stock, staffing and capacity forecasts at the granularity you actually order or roster at, with seasonality and promotions handled.
A model that is accurate in a notebook and absent from the workflow has changed nothing. We build models that land where the decision is actually made — in the ordering screen, the queue, the pricing tool — and that keep working after the data drifts.
Gradient boosting on well-prepared tabular data continues to beat more fashionable approaches for most business prediction problems. It is cheaper to run, easier to explain and simpler to monitor. We use language models where language is involved and classical methods everywhere else.
The largest determinant of model quality, and the part that requires understanding your business rather than the algorithm.
Time-based splits for anything temporal, because random splits leak the future and flatter the model.
Feature attribution so a person can see why a prediction was made — usually a requirement, not a nicety.
Input distributions and accuracy tracked in production, with alerts when reality moves away from the training data.
Stock, staffing and capacity forecasts at the granularity you actually order or roster at, with seasonality and promotions handled.
Which customers or accounts are at risk, with the contributing factors surfaced so the score is actionable.
Product, content and next-best-action recommendations that respect stock, margin and business rules.
Automatic categorisation of tickets, transactions and documents, with confidence thresholds that route the uncertain cases to people.
Unusual transactions, sensor readings or usage patterns flagged for review, tuned so alert volume stays workable.
Data-driven pricing and allocation within constraints you set, with simulation before anything goes live.
Getting a model into production and keeping it healthy is a bigger engineering task than training it. We build the pipeline first and treat the model as a component that will be replaced repeatedly.
The same transformations at training and inference time, so a model does not silently degrade from a mismatch.
Reproducible, versioned runs with tracked data, parameters and metrics.
Batch scoring or a real-time endpoint, whichever the decision actually needs — real time is often unnecessary.
Drift, accuracy and prediction distribution watched continuously, with retraining triggered by evidence.
Capability in isolation changes nothing. These are the systems this normally has to reach in order to be useful.
Training data read from the warehouse, with predictions written back for reporting.
Scores delivered into the ERP, CRM or ordering screen where the decision is made.
Forecasts and scores available in the dashboards leadership already reviews.
Predictions triggering workflows — a reorder, an alert, a review task.
A forecasting model in operation, from warehouse data to an order recommendation someone can act on.
The same six stages apply to every AI build we run. The evaluation set is created before development starts — without it there is no way to tell an improvement from a regression.
Is a model the right instrument here? A deterministic rule is cheaper, faster and easier to audit when the problem allows one. We say so when it does.
Sources identified, access agreed, content parsed and indexed. Data quality sets the ceiling on everything after this point.
Real examples with known-good outcomes, collected before any development. This is what makes later decisions measurable.
A working slice scored against the evaluation set — accuracy, latency and cost per request all measured rather than estimated.
Wired into the real system with permissions, rate limits, spend caps, fallbacks and human approval wherever the action is irreversible.
Quality, cost and drift tracked after launch, with the evaluation set re-run whenever a prompt, model or data source changes.
These are not add-ons quoted separately. They are the conditions under which a system like this is allowed near production data.
Feature attribution retained per prediction, so an affected person can be given a real reason.
Performance measured across relevant groups, not just in aggregate where disparities hide.
Only the features that genuinely improve the model, which is usually far fewer than the ones available.
Staff can override a prediction, and those overrides are captured as training signal.
Forecast demand per SKU and location so reorder points reflect the pattern rather than a fixed rule set years ago.
Score transactions for fraud risk in real time, with explanations attached for the analyst who reviews them.
Predict appointment demand and non-attendance to schedule staff and reduce wasted slots.
Comparable-based valuation estimates as an input to professional judgement, never as a replacement for it.
Flag projects trending toward overrun early, based on how comparable projects historically behaved.
Anticipate contact volume by channel and topic so staffing matches the coming week rather than the last one.
For most tabular business problems, a few thousand labelled examples with genuine signal is a workable starting point — and two years of history for anything seasonal. Data quality and relevance matter far more than volume. We assess what you have before proposing an approach, and sometimes conclude the honest answer is to fix collection first.
Classical machine learning for numbers, categories and time series; language models for text and unstructured content. Gradient boosting still outperforms neural approaches on most tabular business data, and it is cheaper to run, faster to score and easier to explain. Using a language model for a forecasting problem is a common and expensive mistake.
Because we monitor it. Input distributions and prediction accuracy are tracked continuously, and drift triggers an alert and a retraining decision. Models degrade quietly as the world changes — a model nobody is watching is an unmeasured liability, which is why monitoring ships with the model rather than after it.
Yes, and we regard this as a requirement rather than an enhancement. Feature attribution is stored with each prediction and surfaced in the interface. People will not act on a score they cannot interrogate, and in regulated contexts you may be obliged to explain it.
Tell us what it is and what data surrounds it. We will assess whether a model would genuinely beat the current rule.