AI Architect · Financial Services · London
I design and build production AI systems for regulated financial services — and ship my own.
25+ years across banking and insurance, a research background in biomedical engineering, and a habit of writing the code — not just the strategy. RAG, agentic systems and MLOps on AWS and Azure.
How I work
Three things a client is really buying: the design, the working code, and the judgement that comes from running my own AI systems in production.
Twenty-five years defining enterprise AI, cloud and data platforms for banks, insurers and government. I know what a client's risk, security and audit functions will actually sign off — and design so they do.
Hands-on RAG and multi-agent systems in Python and Rust — FastAPI, LangGraph, vector search, LightGBM, MLOps pipelines — deployed as real, monitored services rather than proofs of concept.
Three AI systems I designed and built run on my own infrastructure — live on the web and open on GitHub: cloud-cost intelligence, quantitative ML, and linguistic bias detection.
Selected builds — my own systems
These aren't slideware. Each is a system I architected and wrote end to end, running in production on my own infrastructure. The point of each card is what it proves I can build for a client.
A Claude-powered agent with nine specialised tools that interprets multi-cloud billing and reasons over 56 cost-optimisation rules across AWS, Azure and GCP — then generates Terraform remediation and ranked recommendations. Rust/Axum backend, React 19 front end, PostgreSQL, Redis and a Qdrant vector store for retrieval.
Proves: agentic GenAI over a real domain, full-stack product engineering, and RAG retrieval — in my strongest specialism.
A Ridge / LightGBM / GRU ensemble with a custom directional-penalty loss, generating signals across 197 assets in equities, FX and crypto, validated with walk-forward backtesting. Built entirely in Rust as systemd microservices over PostgreSQL, exposing a signals API and an agentic reasoning service, tracking a live paper portfolio from a $100k base.
Proves: a full ML lifecycle in production plus disciplined quantitative validation and deployment.
An LLM-driven engine that analyses framing and bias across sources, combining retrieval over a vector store with structured reasoning to surface the competing narratives inside a body of text.
Proves: applied NLP turned into a live product, and Responsible-AI thinking on bias and transparency.
Reference architectures
For each major engagement, an open architecture write-up and a working reference implementation — the 2026 design, the trade-offs, the measured numbers, and the code on GitHub.
Ten acquired insurers, ten ways of spelling 'policy'. The mapping was six weeks of human archaeology per company. Here's the 2026 version — where a model proposes the mapping, a human approves it, and nothing leaves the building.
↗ Read the architectureEvery regulated organisation is asking the same question: how do we get AI without our data leaving the building? Most answers guard the boundary in a system prompt. That isn't a control — it's a request.
↗ Read the architectureTwo million customers, points that had to appear instantly, a three-day batch cycle, and a 99.9% SLA. No LLM anywhere near it — which is rather the point.
↗ Read the architectureAn assistant that answers an agent's question mid-call, in under two seconds, with a citation — and says nothing at all when it isn't sure. In a bank, the refusal is the feature.
↗ Read the architectureEnterprise AI delivery
A selection of enterprise work — chosen for relevance to financial-services and insurance AI. The domain knowledge behind these is the part most engineers can't fake.
Background
An engineering and research foundation, kept current with hands-on ML certification.
Contact
Available for select AI architecture and engineering contract engagements in financial services. If you need someone who can own the design and write the code, get in touch.