Description

I make AI systems work in production, and keep them working. I also build the product around them.

Most AI projects stall in the same place: a prototype that demos well but nobody trusts enough to put in front of real users. Closing that gap is what I do. I spent the last year leading the team behind Cambridge's automated marking APIs, the services that mark written and spoken English exams, where other teams' models had to stay dependable in front of real exam candidates. The platform held 99.98% uptime across twelve months.

On my own products I do the model side as well: fine-tuning transformers on serverless GPUs, calibration, and the evaluation that decides whether generated output is fit to publish. Frozen benchmarks, held-out folds, fail-closed quality gates. If the checks fail, nothing ships.

Those products are web apps, and I build the whole of them. React and Next.js on the front, Hono and Workers behind, D1 and R2 for state, and the model services underneath. If you need someone who can take a thing from an idea to something people can use, that's the job I do most often.

What clients usually bring me in for:

Taking an LLM or agent prototype into production and making it reliable.

Building the platform around models someone else trained: pipelines, inference, monitoring, cost control.

Full-stack product work in TypeScript, with or without the AI.

Proving whether a model is good enough, with evaluation that stands up to scrutiny.

Leading a small team, or working alongside one.

I work remotely from London through my own company. TypeScript and Clojure by preference, Python where the machine learning lives, AWS and Cloudflare underneath. 30 years as an engineer, most of it hands-on.

I direct AI coding agents the same way I direct a team: decompose the work, verify the output, stay accountable for the result.

My products, writing and code are at tre.systems, with most of the source public at github.com/tre-systems.

Industry field of expertise

Languages

English
Native or bilingual

Workplace preferences

Remote only

Primarily works remotely

Cambridge University Press & Assessment
API Team Lead
EDUCATION AND E-LEARNING
April 2025 - May 2026 (1 year and 1 month)
London, United Kingdom
I led the team behind Cambridge's automated marking APIs, the Text and Speech services that mark written and spoken English exams, including Write & Improve and Speak & Improve. My job was making other people's machine learning dependable in front of real exam candidates. The platform held 99.98% uptime across twelve months at around a 1% change-failure rate.

I stayed hands-on throughout. I built the Clojure and AWS orchestration service that drives text assessment, productionised Python inference as Dockerised SQS workers with retries, idempotency and alarms, and was one of the most active contributors to the Terraform monorepo underneath. I also ran the response-validity suite that decides whether a submission can be marked at all: off-topic, gibberish, offensive and copy detection. The detectors were the ML teams' work; running them dependably was mine.

Alongside the engineering I moved the team to continuous delivery, owned our part of Cyber Essentials Plus, and built the documentation culture: architecture as code, RFCs, runbooks and a public API docs site.
Clojure Amazon Web Services Python Terraform
Total Reality Engineering
Independent AI Engineer
EDUCATION AND E-LEARNING
May 2026 - Today (3 months)
London, United Kingdom
I build and run live AI products for language learning and assessment, on Cloudflare Workers with serverless GPU inference on Modal. That means the whole stack: fine-tuning and deploying transformer models, post-hoc calibration, grammar correction alongside the scorer, and the infrastructure underneath.

I gate models on measured evidence rather than impressions: frozen benchmarks with held-out folds, fail-closed quality checks on generated content, and side-by-side comparisons scored on task success, convergence and failure modes. If a quality gate fails, nothing ships.

I direct AI coding agents the same way I direct a team. Decompose the work, verify the output, stay accountable for the result.
Typescript Cloudflare Pytorch Python
English Language iTutoring (ELiT)
Senior Software Engineer
EDUCATION AND E-LEARNING
February 2023 - April 2025 (2 years and 2 months)
London, United Kingdom
Senior engineer on Speak & Improve, an AI feedback tool that helps English learners worldwide improve their speaking, built in Clojure around custom marking models. I later moved to the API team, which became TAPI within ARC once ELiT was merged into Cambridge in 2025.
Clojure Machine learning Amazon Web Services