Associate Director, Global AI & Data Science · Kyndryl · London
I build and govern agentic AI: an enterprise agent framework and platform built from scratch, and policy-as-code guardrails that turn business rules into enforceable decisions. Before that, I fine-tuned multimodal LLMs for document AI and computer vision in FinTech (full-parameter, LoRA and QLoRA), and built NLP and RAG systems in R&D, on a foundation of a PhD and lecturing engineering at the University of Edinburgh.
About
I lead AI engineering work at Kyndryl, where I've built core pieces of an enterprise agentic framework in pure Python (function calling, orchestration, MCP servers with a registry and gateway, agent memory and full observability), and now lead Policy as Code: turning rules extracted from business documents into Rego policies that Open Policy Agent enforces on every agent decision.
Before Kyndryl I trained and productionised multimodal LLMs and computer-vision models for document AI in trade-finance compliance at Traydstream, using full-parameter, LoRA and QLoRA fine-tuning, and built NLP, LLM and RAG systems at Twyn. My background is scientific: a PhD in materials science, a first-class BSc, and a year lecturing engineering at the University of Edinburgh, which is where I learnt to care about rigorous evaluation and well-made tools.
Projects
Enterprise AI at Kyndryl, a retrieval system over hundreds of hours of video, and side projects and open-source contributions.
Ask a public figure a question and the app plays the exact moment in a real interview where they answer it. Retrieval-augmented generation over 371 videos (284 hours) in English and Greek, with an evaluation harness behind every change: clips now start within a second 99% of the time, and searches take about 1.3 s.
# illustrative example package agent.guardrails default allow := false allow if { input.tool in data.client.approved_tools not exceeds_spend_limit }
Part of Kyndryl's enterprise governance capability for agentic AI: policies extracted from business documents are converted into enforceable Rego rules and evaluated by Open Policy Agent, integrated into customer environments to govern agent decisions and tool execution with auditable, customer-specific guardrails.
Core components of a pure-Python agentic framework: function calling, MCP servers with an MCP Registry and Gateway, orchestration, tool discovery and execution, agent memory and concurrency, with observability through OpenTelemetry, OpenLLMetry, Prometheus, Grafana Mimir and dashboards.
A web application that uses Google Custom Search and OpenAI's GPT to research a topic, write it up in LaTeX and compile it into a downloadable PDF, so anyone can produce professionally typeset documents.
A productivity tool that captures any part of the screen, extracts the text with OCR and sends it to an LLM API through keyboard-driven automation, showing the answer in a Tkinter window with code highlighting. Image description uses the vision model.
Contributions to Spartan, an open-source game engine (3k+ stars) focused on architectural quality and performance: physics for flotation and movement dynamics that make its worlds more realistic and interactive.
Classification, regression and time-series forecasting with XGBoost and decision trees, plus neural networks and convolutional networks applied to real problems, from house prices to ovarian cancer image classification.
Experience
Leading delivery and client integration of Policy as Code: rules from business documents compiled to Rego and enforced through Open Policy Agent, deployed with Terraform and Argo CD.
Built an enterprise agentic framework from scratch in pure Python, with MCP servers, registry and gateway, and end-to-end observability.
Trained and productionised text and multimodal LLMs for FinTech (full fine-tuning, LoRA, QLoRA), object-detection pipelines and agentic compliance checks.
LLMs in production on Azure: prompt engineering for multi-label classification, fine-tuning, multi-agent systems and RAG pipelines.
Lectured, designed and examined Environmental and Water Engineering; taught Engineering Mathematics and Design.
Thesis: design and synthesis of porous adsorbents for water purification.
Writing
A case study in retrieval-augmented generation over 284 hours of video, and the evaluation work that made it accurate and fast.
Why LaTeX beats Word for serious documents, and the tool I'm building to make it accessible.
The expected number of throws of n dice until one exceeds r, derived and checked with a Monte Carlo simulation.
Contact
Open to conversations about data science, LLM applications and research collaborations. Email is the quickest way to reach me.