Outcome 02
Turn AI into production software
Most AI projects die in the gap between a convincing demo and a system you can depend on. I’ve spent years on the production side of that gap.
The symptoms
- You have a promising prototype but no path to production.
- There’s no honest way to tell whether the model is actually good.
- Data is scattered and the pipeline breaks quietly.
- Nobody’s sure what it will cost — or whether it’s safe to ship.
What I do
- 01Design the data platform underneath the model — streaming, storage, and recovery that don’t lose data.
- 02Build LLM and MCP systems that put live business data safely in front of the people who need it.
- 03Put evaluation and A/B testing in place so model quality is measured, not assumed.
- 04Own the path to production: deployment pipelines, monitoring, and cost.
Proof
Co-founder & CTO — ML personalisation
Designed end-to-end model deployment and continuous A/B testing, serving real-time recommendations to millions of users for major broadcasters — 10,000+ requests per second at a 19 ms p99, under contractual SLAs.
20%+ lift in pageviews
AI/ML platform engineer — e-commerce group
Built a ClickHouse lakehouse fed by real-time streams from 13+ webshops, and an MCP server giving non-technical staff LLM-powered analytics over live data through Claude.
Live analytics for non-technical users
Lead data engineer — broadcaster
Architected a next-generation analytics and ML platform combining batch and real-time pipelines for recommendations and audience insight.
50+ TB processed per day