Binary and multi-class classification models for problems with clear outcome categories: fraud or legitimate, churn or retained, approved or declined, high-priority or routine. Gradient boosting models (XGBoost, LightGBM, CatBoost) handle tabular data with mixed feature types and produce probability outputs that rank and score rather than deliver hard binary decisions -- a fraud score of 0.87 is more useful than a binary flag. Logistic regression baselines provide a calibration reference before moving to ensemble methods. Feature importance via SHAP values gives regulators and internal teams the explainability evidence they need for model sign-off. Models trained on your labelled historical data, evaluated on stratified held-out sets with precision-recall curves, and deployed behind FastAPI or BentoML with defined performance SLAs.

Time-series forecasting models for demand, sales volume, resource usage, and capacity planning -- the decisions where being wrong by 20% has direct inventory or staffing cost. Prophet handles strong seasonality and holiday effects with minimal configuration; SARIMA and SARIMAX perform well on stable, lower-frequency series; LSTM and Temporal Fusion Transformer networks capture complex long-range dependencies in high-frequency data like hourly energy demand. Exogenous variables (promotions, weather, competitor pricing, macroeconomic signals) fed as regressors improve accuracy for categories that respond to external events. Multi-horizon forecasts delivered at product, SKU, location, or segment level with prediction intervals so planners know the confidence band, not just the point estimate. Output integrated into your ERP (SAP, NetSuite), inventory system, or planning tool via API or scheduled data push so forecasts drive procurement decisions rather than sitting in a report.

Churn prediction models that score every customer on their likelihood of leaving before they do -- delivering your retention team a ranked list updated weekly so intervention effort goes to the accounts most likely to respond. Feature engineering draws from product usage logs (login recency and frequency trends, feature breadth decline, session depth), support ticket history (volume spikes and unresolved escalations), billing signals (payment delays, downgrade attempts), and engagement data (email open rate decline, NPS trajectory). XGBoost and LightGBM handle the mix of numerical, categorical, and behavioral features typical in SaaS churn datasets; class imbalance is handled via stratified sampling and cost-sensitive learning rather than oversampling artifacts. Threshold tuning is calibrated to your intervention budget -- if your CSM team can action 200 accounts per week, we optimise precision at that recall level, not the headline AUC. Automated retraining pipelines triggered by data drift signals (monitored via Evidently AI) keep the model accurate as product behavior evolves over quarters.

Statistical and ML-based anomaly detection for fraud, equipment failure prediction, network security, and manufacturing process quality control -- domains where false negatives are costly but false positives destroy operational trust. Supervised detection (XGBoost, Random Forest) performs best when you have labelled historical anomalies and the anomaly patterns are stable; unsupervised methods (Isolation Forest, Local Outlier Factor, Autoencoder neural networks) handle cases where anomalies are rare, novel, or unlabelled. For time-series anomalies in sensor or transaction data, LSTM-based sequence models detect contextual anomalies -- events that are individually normal but statistically unusual given prior context (a login at 3am from a new geography immediately before a large transaction). Real-time scoring via Kafka stream processing for sub-second fraud detection; batch pipelines for equipment sensor review and overnight audit queues. Precision-recall thresholds tuned to your false positive tolerance -- lowering the threshold catches more anomalies but increases the review queue; we scope this tradeoff during discovery rather than delivering a model with a single arbitrary cutoff.

Collaborative filtering, content-based, and hybrid recommendation models for product discovery, content personalisation, next-best-action, and cross-sell -- trained on your interaction history rather than generic behavioral benchmarks. Matrix factorization (ALS, BPR) and neural collaborative filtering (NCF) learn from co-purchase and co-engagement patterns when sufficient user interaction data exists. Content-based models using TF-IDF or sentence embeddings handle cold-start situations where new users or new catalog items have no interaction history yet. Two-tower neural networks combine both signals for platforms where both content metadata and behavioral data are rich. Candidate generation (approximate nearest neighbor search via FAISS or Pinecone) followed by a reranking layer (LambdaMART or a learned ranker) separates the retrieval and ranking stages for low-latency production serving. A/B testing with holdout groups validates lift in click-through, conversion, or session depth against a non-personalized baseline before full rollout -- preventing the false precision of models that look good on offline NDCG metrics but don't improve business outcomes.

Production deployment infrastructure for ML models: model serving via FastAPI or BentoML with versioned endpoints, blue-green deployment for zero-downtime model updates, and rollback capability when a new model version underperforms in production. MLflow or DVC tracks experiments, datasets, and model versions so every production model has a complete lineage record -- what data it was trained on, which hyperparameters, and what evaluation metrics were achieved. Feature stores (Feast, Tecton) ensure that the same feature computation logic runs identically during training and serving, eliminating training-serving skew -- one of the most common causes of production ML underperformance. Data drift monitoring via Evidently AI or Arize watches input feature distributions and alerts when they diverge significantly from the training distribution, and model performance monitoring tracks prediction accuracy against ground truth labels as they accumulate. Automated retraining pipelines (Airflow or Prefect) trigger when drift signals cross defined thresholds -- retraining on a schedule without drift evidence is wasteful, but waiting for drift to compound into visible accuracy loss is expensive.