Machine Learning
Machine learning is not a single technique, it is a broad family of approaches for extracting patterns from data and using them to make predictions, decisions, or generate content. We apply the right ML methods for each problem: from simple, interpretable models to state-of-the-art deep learning architectures.
ML Techniques We Work With
- Supervised Learning, Classification and regression with gradient-boosted trees (XGBoost, LightGBM), random forests, and linear models for tabular data. Deep learning (CNNs, transformers) for images, text, and time series.
- Unsupervised Learning, Clustering (K-Means, DBSCAN, hierarchical), dimensionality reduction (PCA, UMAP, t-SNE), and anomaly detection for unlabelled datasets.
- Natural Language Processing, Text classification, named entity recognition, sentiment analysis, question answering, summarisation, and retrieval-augmented generation (RAG) with large language models.
- Computer Vision, Object detection (YOLO, Detectron2), image segmentation, image classification, and optical character recognition for document processing and industrial inspection.
- Time Series Forecasting, Demand forecasting, anomaly detection, and trend analysis using classical (ARIMA, Prophet) and neural (N-BEATS, TFT) methods.
- Recommendation Systems, Collaborative filtering, content-based filtering, and hybrid systems for personalising product, content, and service recommendations.
Our ML Toolchain
We use Scikit-learn for classical ML, PyTorch for deep learning, Hugging Face Transformers for foundation models, and Optuna for hyperparameter optimisation. Experiments are tracked in MLflow or Weights & Biases, datasets are versioned with DVC, and models are deployed via ONNX Runtime or TorchServe for maximum portability and performance.
Responsible ML
We document every production model with a model card covering training data provenance, evaluation metrics, known failure modes, and intended use cases. We run bias audits using tools like Fairlearn and provide SHAP or LIME explanations for models that affect significant decisions.