Promoted to Staff Engineer to build the platform for AI development across ZOE.

• Built llmrouter, an open-source latency-aware sidecar that routes LLM requests to the fastest provider, needing no application changes (injected via a Kyverno mutating webhook).
• Forecasted required provisioned LLM throughput, collaborating with Growth and Analytics teams.
• Led fine-tuning and eval of open weight models to produce and serve proprietary AI models.
• Implemented observability and alerting for AI features, monitoring for drift and throughput limits.
• Led migration to centralised secrets injection (Secrets Manager) across ZOE’s kubernetes fleet, backwards-compatible rollout with no disruption to running services.
• Established best practices and tooling for LLM development across the business.
• Mentoring and hiring up to Lead AI engineer.

Shipped AI based photo logging for tracking meals, becoming Zoe’s key product feature.

• Implemented with multimodal LLMs and low latency two-stage semantic search (RAG).
• Built evaluation framework to measure AI performance to a rigorous standard with ML metrics.
• Architecture design, iteration and load testing to support 10x scale-up and low latency AI.

Developed and deployed realtime product ranking and recommender models to the Bloom & Wild website and app, delivering substantial uplifts in revenue.

• Built out serverless ML and data infrastructure on AWS using terraform.
• Identified AWS savings totalling 25% of the business’ AWS spend.
• Improved API performance, e.g. reducing AWS Lambda API p95 latency by 10x.
• Introduced and deployed MLOps tooling such as mlflow and Feast.
• Raised engineering standards, e.g. use of mypy, CI pipelines and a python community of practice.
• Line management, introducing agile working practices and mentoring in ML techniques.

Led a cross-disciplinary squad of engineers and data scientists.

• Developed and deployed ML models to predict retention probability, detect anomalous trading.
• Trained XGBoost models to forecast flight price changes and highlight deals to customers.
• Developed new feature flagging and experimentation platform used across the entire tech stack.
• Introduced sagemaker, dask, databricks and mlflow as ML platform for the data science discipline.
• Hired, developed and promoted data scientists up to principal level.

As the first data scientist hire of Vodafone UK, I helped scale up the function, training and deploying the first models to production and delivering significant uplifts in personalised CRM campaigns.

• Developed and deployed collaborative filtering recommender system with pyspark for personalised marketing.
• Trained models to determine ‘Next Best Action’ to support agents with customer calls.
• Led interviewing and training of Data Science hires in Vodafone UK.

Trained and deployed models for music machine learning and recommendation, e.g. building a voice classifier and collaborative filtering music recommender system for major music label using their large corpus of audio data.

Collaborated with product team to develop novel algorithms for personalising music recommender systems. Designed and prototyped recommender system concepts that were incorporated into the Beosound Moment product.