Production AI,

Decoder-only transformers trained on next-token prediction. State of the art for open-ended generation, reasoning, and instruction following.

Bidirectional transformers that produce rich contextual embeddings. Best suited for classification, NER, and semantic similarity.

Seq2seq architecture mapping input sequences to output sequences. Used for summarisation, translation, and question answering.

Generative models that learn to reverse a noise process. State of the art for high-fidelity image and media synthesis.

Region-based and anchor-free detectors for localising and classifying objects in images. Applied in insurance claim assessment and industrial inspection.

Hierarchical spatial feature extractors. Backbone of most computer vision pipelines for classification and representation learning.

Recurrent architectures that model temporal and sequential dependencies. Applied in OCR, time-series analysis, and sequence labelling.

Attention-based architecture that unified NLP and vision. Trained via self-supervised objectives before task-specific fine-tuning.

Adversarial generator–discriminator framework for producing realistic synthetic data and high-quality image generation.

Message-passing networks that learn on graph-structured data — knowledge graphs, recommendation systems, and molecular modelling.

Policy optimisation through environment interaction. Underpins LLM alignment (RLHF) and sequential decision-making agents.

Compressing large teacher models into faster, smaller student models while preserving accuracy — critical for production deployment.

Classical algorithms for structured and tabular data. Fast to train, interpretable, and often the right tool before reaching for deep learning.

Hands-on experience with core AWS services for data engineering, ML workloads, and application infrastructure.

Deep production experience across the Azure ecosystem — from ML pipelines and vector search to DevOps and real-time compute.