Selected work, newest first.
Stream: a data function built for AI agents
I lead data at Stream (formerly Wagestream), a financial wellness fintech. I report to the CPTO.
People and AI agents both ask data questions across the company, and both need the same definitions. So I built an MCP-based context layer. It gives AI agents governed access to metric definitions, schema documentation and lineage. Self-serve analytics and agentic coding now run on trusted definitions instead of guesswork.
I also launched an in-house, AI-first BI platform that replaced the incumbent commercial tool. It is part of a data-stack cost programme on track to save ~$150k/year.
From Stream’s proprietary data, I created real-time UK macroeconomic indicators, which we shared with senior external stakeholders.
The team is hub-and-spoke: a lean core, plus analytics engineers embedded in Operations, Credit and the US.
Tarabut: an open banking data platform from scratch
I was the first data hire at Tarabut, MENA’s leading open banking platform. I designed and built the enterprise data platform on AWS, GCP, Snowflake and dbt. After M&A, I led the integration that put UK and Middle East data on one platform.
On that platform we delivered ML-powered bank transaction enrichment services across 6 regions, and external data products that opened new revenue streams.
The team grew from 1 to 10 people across 4 countries. I also chaired the Data Governance Committee and set one governance framework for the whole group.
TrueLayer: the company’s first ML products
I was the joint-first data hire (~employee #20) at TrueLayer, the UK’s first open banking unicorn. I built its first ML services, which turned raw bank transactions into financial insights:
- Transaction categorisation for spend analysis
- Income estimation API, shipped as a commercial product
- Affordability signals: rent detection, recurring payments, and spending patterns
The services ran at 99.9% availability with weekly model updates. They became TrueLayer’s first revenue-generating data products beyond the core API.
Earlier work
- Black Swan Data: led data science for a platform that predicted consumer trends from 100M+ social media posts a month, for Fortune 500 clients.
- IBM: built ML solutions for customer analytics and fraud prevention in Tier 1 banks.
- PGS: used ML to automate quality control of seismic swell noise filtering, which cut manual QC time by 70%. Presented at EAGE London 2013.
On the side
FPL agent. I built an AI agent that manages my Fantasy Premier League team. An ML model predicts player points, an optimiser picks the squad within the game’s constraints, and an LLM agent with tools makes the weekly decisions. I wrote it up in three posts:
- Engineering an FPL Brain
- Predicting Haulers, Not Averages: An ML Pipeline for Fantasy Football
- Building an FPL Agent: Tools, Context, and a Checker
AI Doom Boom Bias Tracker (2024). An NLP pipeline on GCP that classified how news outlets frame AI. Write-up ・ GitHub