What I've built.

A job search tool that isn't allowed to make things up about you.

Plenty of AI CV tools will happily invent a promotion if it improves the match. OutRung is built around not doing that. You upload your old CVs once, and it turns the evidence in them into a single profile. From there it finds roles, scores how well each one fits, and writes a CV for the ones worth applying to. The model can choose, cut and rephrase, but it can't add anything the profile doesn't back up, and that rule is enforced by the pipeline, not left to a prompt. I'm the only engineer, so I design, build and run all of it, which keeps me honest about what production AI really takes: evals, tracing, alerts and a cloud bill. It's backed by Microsoft for Startups.

• Grounded CV pipelineA multi-stage LLM pipeline on Azure AI Foundry that parses jobs, scores fit and writes CVs, with guardrails that tie every line back to the profile.
• Evals from real failuresAround ten thousand production LLM calls traced. When one goes wrong, the anonymised case becomes a regression test.
• The production estateFastAPI, PostgreSQL and Docker on Azure, all in Terraform, with CI/CD, alerts, cost budgets, backups, and the database reachable only over Tailscale.
• Agentic engineeringBuilt with Claude Code and Codex inside harnesses I wrote: repo-specific skills, guard hooks, multi-agent PR review, and a human approval gate before anything ships.

• 40+users who found it on their own, through search and LinkedIn
• ~10kproduction LLM calls traced
• 1engineer: me
• 0achievements invented by the AI

• LLM pipelines
• Evals
• Azure AI Foundry
• FastAPI
• PostgreSQL
• Terraform
• Docker
• Claude Code

Teaching a 3D-printed microscope to count bacteria.

The standard test for bacteria in drinking water goes like this: send a sample to a lab, wait a day for colonies to grow, then have a trained technician count them by eye. The places that most need the answer are usually the furthest from a lab. WaterScope started in 2015 with a handful of Cambridge researchers, biologists, physicists and engineers, who wanted the whole test to run in a box, on site. The microscope was 3D-printable; I wrote the software around it, from the camera control and the interface on the device to the computer vision and deep-learning models that find and count the colonies on a Raspberry Pi. We took it to field trials in Tanzania, India and Mexico. I was doing a PhD in photonics at the time. This side project is where my AI career started.

• £2Mnon-dilutive funding from EPSRC, BBSRC and Innovate UK
• 10×faster results than existing kits, at under half the hardware cost
• 2017Cambridge Vice-Chancellor's Impact Award, for the team
• 1granted patent, on the test cartridge and sample handling

• Computer vision
• Deep learning
• Edge AI
• Raspberry Pi
• Python
• Flask
• Vue.js

AI for organisations that need it to work, not just to demo.

Softwire builds software for other organisations, and I lead its AI work: deciding with each client what's worth building, then building it with their people. I was the first at Softwire to become an Anthropic Certified Architect. Some of the work is hands-on delivery. Some is helping a whole organisation start using AI safely, from engineers working alongside coding agents to finance and HR teams building their own tools. And some is the discovery and the proposals that come before any of it.

• AI accounting productAn accounting startup wanted AI at the heart of its product, and not much more of a brief. The financial rules stay in deterministic code; LLMs read the messy statements, and every model or prompt change goes through evals. Discovery to production in seven weeks.
• Citizen developersTraining and a low-code programme on Copilot Studio and Power Automate, so HR, finance and marketing teams build their own automations.
• Coding agents for teamsHarness and context engineering for human-in-the-loop coding agents, set up so engineers at every level can build well with them.
• AI on personal dataGetting AI cleared for sensitive personal data: data residency, governance and risk, with business, technology and risk teams agreeing on a way forward.

• 7weeks from discovery to production
• 1stAnthropic Certified Architect at Softwire

• LLMs
• Evals
• Agent harnesses
• Context engineering
• Copilot Studio
• Power Automate
• Human-in-the-loop

AI at a fintech, where every word sent to investors is regulated.

SyndicateRoom is an online platform for investing in early-stage UK startups. My job was to make data and AI useful across the company: products for investors, tools for the team, and the training to go with them. Anything that reaches investors counts as a financial promotion under UK rules, so compliance had to be designed into each system rather than checked at the end. The flagship was the investor reports. Thousands of portfolio updates arrived every month and nobody had time to read them; now each investor gets a short note about the companies they actually hold.

• Investor reportsAn LLM reads the month's portfolio updates and writes each investor a personal digest of what they hold.
• Knowledge platformOne RAG knowledge base built from the website, the podcast, government policy and user data, served over MCP and REST to draft pages, posts, ad copy and investor profiles, with the compliance rules built in.
• Operating dashboardCampaign performance, targets and KPIs in one place, with role-based access, feeding the weekly report and the board's discussions on direction and ad spend.
• Startup scoringBayesian and tree-based models to spot high-potential UK startups.
• AI for the teamn8n and Python workflows, and training for design and marketing on AI tools, MCP servers, prompting and keeping a human in the loop.

• 70%open rate on the AI investor reports, up from under 20%
• 30%click-through on the same emails, against a 5% benchmark

• RAG
• MCP
• LLMs
• n8n
• Python
• Gradient boosting
• Streamlit
• GA4

The colour from my PhD, made by the kilo.

Sparxell spun out of the Cambridge lab where I did my PhD, to make the kind of colour I'd spent four years on: cellulose arranged so finely that it reflects light as colour, with no dye and no plastic. In the lab, we made it droplet by droplet on a microfluidic chip. Luxury brands wanted it by the kilogram, and they would notice if one batch came out a shade off. As Head of R&D I hired and led a team of six scientists, and we ran the scale-up like one long experiment: decide what to measure on every batch, read the data, change one thing, repeat.

• Lab to pilot plantDecision matrices, batch-by-batch characterisation and tight feedback loops, until the process held up at pilot scale.
• Luxury brand projectsMore than ten paid innovation projects, worth £250k, for houses including LVMH, Moncler and Gucci. Every milestone and budget met.
• Funding & IPTechnical lead on £3M of grant funding, including an EIC Accelerator and four Innovate UK grants, and three patent applications in fashion and cosmetics. The technology I validated underpinned £5M of seed and Series A investment.

• 1000×scale-up, from grams to kilograms a day
• 3×product performance
• −90%material cost
• £3Mnon-dilutive funding won

• Process scale-up
• Materials characterisation
• Experiment design
• Team leadership

Three years of other companies' hardest problems.

Companies go to TTP when the product they want doesn't exist yet and their own engineers are stuck. That means a new hard problem every few months, with someone else's money and deadline attached. I was key account holder for long-running clients in medical devices, life sciences, advanced materials and manufacturing. I ran the discovery workshops, owned the budgets and the risk, and explained the technical trade-offs to the executives paying for them, while still doing a fair share of the engineering. A lot of the briefs sounded impossible. The job was to make them possible anyway.

• Dispensing robot visionThe eyes and decision logic for a robot that dispenses medicines. A robot handing out medicine can't guess, so every uncertain reading had to fail safe, within what the hardware could do.
• PFAS-free filtrationOil-repelling filter materials without the 'forever chemicals' that usually do the repelling. Two granted patents.
• Microfluidics & sensingMicrodroplet systems, and acoustic devices that move nanoparticles around and sense them.

• £3Mof development projects run as key account holder
• 2granted patents in filtration and microstructured materials

• Python
• OpenCV
• OCR
• Microfluidics
• Microfabrication
• Metrology

Stops your AI agent paying Opus prices to say 'hi'.

Agent sessions run for hours, and every message, even a 'thanks', goes to whichever model the agent is set to, usually the most expensive one. LazyRouter sits in between. A cheap, fast model reads each request first and hands it to the best-value model that can handle it, ranked by price and ELO rating. You list your models in a YAML file and ask for model: 'auto'. Behind that one OpenAI-compatible API sit OpenAI, Anthropic and Gemini; when one is rate-limited it falls back to a similar model, and it compacts long histories to keep the prompt cache warm. OpenClaw users run it for their agents. Around the same time I wrote the OpenClaw integration for Hindsight, an open-source memory system that lets agents remember things between sessions.

• 3providers behind one API: OpenAI, Anthropic, Gemini
• 1setting to turn it on: model: 'auto'

• Python
• FastAPI
• LiteLLM
• LLM routing
• Prompt caching
• Agent memory

An experiment that had to run itself, twenty seconds at a time.

In 2018 I flew on the European Space Agency's 70th parabolic flight campaign, out of Bordeaux. The plane climbs steeply, eases off the throttle and arcs over the top, and for about twenty seconds everything on board floats. Then it pulls out and does it again, thirty-odd times a flight. We were studying emulsions, droplets of one liquid held in another, and how they behave when gravity stops pulling them apart. I designed and built the rig: Arduino and Raspberry Pi control, cameras analysed with OpenCV, and a web interface to drive it. Twenty seconds isn't long enough to fix anything by hand, so the rig had to get it right on its own, every parabola. It's still the best lesson I've had in designing for the moment nobody can step in.

• ~20 sof weightlessness per parabola
• 31parabolas a flight
• 70thESA parabolic flight campaign

• Arduino
• Raspberry Pi
• OpenCV
• Flask
• Vue.js
• Fluid dynamics