Episode #28
The Complete MLOps Lifecycle: From Data to Deployment | Phillip Mortimer
90% of ML projects never make it to production. That's not a talent problem — it's an MLOps problem. Phillip Mortimer is a computer scientist, ex-Chief Scientist and CTO at a London fintech, and one of the only people teaching MLOps at university level (Dauphine University, Paris — 5 years running). In this episode, Phillip walks through the complete MLOps lifecycle: • Data preparation — why EDA is the most forgotten step, and why data pipelines still matter in the LLM era • Model building — Karpathy's 5-stage training cookbook: become one with your data → fit a baseline → overfit → regularise → squeeze out the juice • Experiment tracking — MLflow, Weights & Biases, model registries, and model cards • Deployment — real-time vs batch, Docker containers, inference optimisation with ONNX, vLLM, and TensorRT • Monitoring — data drift, feedback loops, and keeping models relevant • The future — why MLOps is shifting to AI engineering, and why agentic AI is the real breakthrough Key stat: 90% of ML infrastructure cost is inference, not training. If you're not optimising your serving layer, you're burning money every day.
