Machine learning and forecasting
Custom machine learning, measured by what it changes
Forecasting, risk, anomaly detection and recommendation models built against your data, validated on held-out periods and shipped with monitoring, not left in a notebook.
Technology we use for this work
The problem
Accuracy in a notebook is not a business result
Many models are judged on a test score and then quietly abandoned, because nobody connected the score to a decision, or because the data feeding them changed and nobody noticed.
We start from the decision the model supports, agree how success is measured, and ship the monitoring and retraining that keep it honest.
What we build
What does our machine learning and forecasting service include?
Forecasting, risk, churn, anomaly detection and recommendation models, shipped with monitoring.
Forecasting
Time-series forecasting for demand, claims cost, revenue and capacity, with honest intervals and back-testing on held-out periods.
Risk and propensity models
Readmission and deterioration risk, credit and fraud risk, and churn models, with explainability suitable for a risk committee.
Anomaly detection
Detection on transactions, telemetry and operational metrics, with alerts routed to the person who can act.
Recommendation engines
Collaborative and content-based filtering combined in two-tower models that adapt to live interaction, and conversational recommenders that explain their choices.
Computer vision
Document OCR, image classification and defect detection, including a CNN image classifier used in production to verify more than 50,000 profiles.
Feature stores
Features defined once in Feast and served the same way to training and to live predictions, so a model in production sees the inputs it was trained on.
Experiment tracking and managed training
Every run, dataset and metric recorded in MLflow or Weights & Biases, with tuning in Optuna and training on AWS SageMaker, Vertex AI or Azure Machine Learning when your cloud calls for it.
Optimization and scheduling
Planning and scheduling problems, from fleet efficiency to resource allocation, backed by an advisory team with published work in optimization.
Proof
Where have we delivered this?
Clinical NLP and risk models for a US healthcare analytics provider
Clinical NLP, ICD-10 and HCC coding support, readmission risk and claims cost forecasting over large volumes of unstructured clinical notes.
Result: 500K patient records processed by clinical NLP
Read the case studyBook recommendations at scale, with a self-hosted LLM
A hybrid recommendation engine over more than 10 million titles and a goal-based recommendation chatbot on a self-hosted Llama 3 model.
Result: 60% higher click-through on recommendations
Read the case studyMulti-agent retail intelligence on a knowledge graph
A multi-agent system that answers business questions, forecasts and draws its own charts by reasoning over a retail property graph.
Result: 4 specialized agents over one governed graph
Read the case studyContext
Industries and related services
Related services
FAQ
Machine learning and forecasting: common questions
What results have your models delivered?
Examples from past work: 60 percent higher click-through on recommendations, 20 percent of at-risk accounts recovered with a churn model at 75 percent AUC, a 50 percent fleet efficiency gain from telemetry analytics, 80 percent claim cost accuracy for small-bill customers and 70 percent AUC on 30-day heart failure readmission risk.
How do you validate a model before it goes live?
On held-out time periods that the model never saw, against a baseline the business already uses, and with the metric agreed before training. We also check behavior across the segments that matter, not only the average.
Can you explain model decisions to regulators or clinicians?
Yes. For risk models we favor approaches that can be explained, report feature contributions for individual predictions, and document the data, assumptions and limits in a model card.
What happens when the data changes?
Models ship with drift monitoring and a retraining schedule, with shadow or canary releases so a new version is compared against the old one before it takes over.
Start with a two to three week assessment.
Send us the problem and the constraints. You get a written architecture, a risk register and a costed roadmap, and you keep them whatever you decide next.