What I've built and run
I build AI systems end to end, ship them solo, and prove the return. Not slides and not prototypes. Live software with real users, a multi-agent AI system doing real work, and savings a client will put their name to. Here's the infrastructure and the receipts.
The builds
Each one is live or has run on real work. I built these myself.
HouseComply live SaaS · built & run solo
A UK property-compliance platform I built and operate single-handed, product and engineering both. It's in real use, not a demo.
- React + TypeScript front end, ~100 serverless API endpoints, a 59-table data layer, 358 automated tests passing (0 failing).
- End-to-end Stripe billing, and a security-hardening pass I did myself: JWT body-to-token binding, a hard CORS allowlist, signed webhooks with idempotency.
- 708 properties and 373 inspections on the live platform. By my own reckoning, a normal team would spend £1.5–2m to build the same thing.
A multi-agent AI system agents · autonomy tiers · guardrails
My own AI workforce: for each job a business needs doing, I build the specialist for it as a dedicated AI worker, then put a fleet of them to work with real safety rails.
- 27 agents defined as machine-readable JSON against a documented schema, with L1/L2/L3 autonomy tiers (read-only, acts, needs human approval) and a queue/worker/dispatcher design for headless runs.
- Built on frontier models (Claude, including Claude Code), with RAG and retrieval, PII handling and a human in the loop by default.
- This is what runs the day-to-day across my own businesses and my clients' systems.
The history research engine retrieval · multilingual · at scale
The Island House project: an AI research system that reads a place's own sources and brings long-lost material back to life, built on a research spine where every claim is linked to the source that backs it.
- By my own account it covered somewhere around 40 million words, across 10 languages and roughly 1,000 years, translating and reading archive documents and unreadable old legislation.
- Done the traditional way it would have taken hundreds of thousands of man-hours and still carried human error on every page. My software has saved that project more than £600k.
- Underpinned by a sources → entities → facts chain, graded findings registers and a deterministic source collector.
Measured client delivery real ROI · owner-confirmed
Through Client Flow I audit a business, build the AI and automation into the systems they already use, and measure what it gives back over 100 days. "If it isn't paying, we stop."
- At a lettings client (Owen & Owen): about £16k a year of admin removed and roughly £30k of potential fines avoided across 113 properties, confirmed directly with the owner.
- Delivery runs on a CRM with an approval gate and the measurement built in. Nothing goes out without a human signing it off.
Data & deployment infrastructure pipelines · serverless · cron
The plumbing underneath it all, built to run unattended and keep itself honest.
- Scheduled serverless collectors that pull from official and open data (HM Land Registry, The Gazette, planning, legislation change-detection) into a clean, deduped data layer on cron.
- Everything deployed and operated by me: Vercel and Cloudflare Workers, DNS and email routing, secrets kept out of the repo by design.
The stack
The person behind it
Self-taught, and all in. I run the AI across several of my own companies myself, max out three AI subscriptions at once, and build faster than I can keep up with — not as a day job, as the thing I actually do. Years of it now, hands on every day. It's saved me and the businesses I work with an enormous amount of time, money and headache along the way.
Before this I built and sold gyms (one past a million in turnover), played international rugby league and won as a British jiu-jitsu champion. Same pattern every time: learn it properly, outwork the room, and stand behind the result.