I lead AI product and deployment: evals, safeguarding and agents in production.
Most of my work is hands-on: product calls, code, evals, and the safeguards around a model that talks to real people. At Zero Gravity I led six product engineers and a designer on an A-Level AI STEM tutor built on Claude, with each engineer owning a problem from prototype to production. I wrote 28% of the commits. It went from a 30-school beta into 850+ UK schools, and the government picked it for its AI Tutoring Tools Pioneers Programme: eight providers chosen from 53 bids.
Everything I build and write lives on ElliotOS, a personal site run like a product: live projects, a changelog kept by agents, and the evals that check them.
What I work on
- AI in production: the Zero Gravity AI STEM tutor. Four subjects, eight exam boards, Socratic by design. Every release passed a golden set of real past papers and official mark schemes: bare model 67%, retrieval 79%, our covered content 99%+. I open-coded production traces into failure modes (sycophancy on "I don't know", inconsistency across modes, hallucinated marks) and turned them into a pass/fail LLM judge on every session
- Safety layers for AI that reaches people who can't afford a wrong answer: ward, a UK safeguarding layer for LLM apps serving under-18s, eval set included
- Prototypes for people who can't afford a wrong answer: lease-citations, where every quote is located in the lease before it renders as evidence, and persona-interviews, interviews with synthetic survey respondents where every claim shows whether it came from the data or the model
- Human judgement inside agent workflows: crux, a field study of what you still decide when the agent writes the code, and dog-years, experiments on why agents estimate in human weeks
- How I split my own work with machines: an open-loop system built on Claude Code that holds capture, reminders and follow-ups so I keep the calls. It publishes a weekly ledger that can show it failing, on /loops
- Claude Code tooling I use myself: Skill Potions (skills and hooks) and an MCP server for the UK home-buying journey, built on live SRA, FCA and Companies House data
- Career ops as software: boulot, the local system that ran my own search. Three adversarial agents argue over every application before it goes out
- Current experiment: duty-desk, a simulated AI duty desk for small-group tour operators, from the crisis ops I ran at Flash Pack. Write-up coming on ElliotOS
- Four times a founding hire. Flash Pack (employee #10, pre-seed to Series A), Farewill (SRA/FCA regulated probate), Zero Gravity (hire #4, Head of Operations & Enterprise Partnerships, then Head of AI Product & Deployment, 2022 to 2026). Co-founded MealsForTheNHS in the pandemic: £1.8m raised, 303k meals in 100 days
- 🤝 Enterprise partners at Zero Gravity: paid programmes inside the early-careers hiring of HSBC, Morgan Stanley, KPMG, Ashurst, Accenture and Snap. We doubled the size of those accounts and got through their due diligence on ISO 27001 and 9001, which I delivered as sole DPO
- 🎙️ On Just Now Possible with Teresa Torres: how we built Zero Gravity's career co-pilot for students from low-opportunity backgrounds
- ✍️ Writing: Shipping AI to users who can't afford bad advice · The product engineer and the end of the handoff · The loop was never the hard part · more on ElliotOS
- BlueDot Impact AI Governance and Alignment. Languages BA (Birmingham and Fudan): French and English native, Spanish intermediate, Chinese HSK 5
🌱 Building an AI product team? I'm looking for a senior or founding AI product seat where I work with customers, build, and own delivery. London or remote from the UK. Say hi on LinkedIn.