Software Engineer · Belfast
I build AI systems end to end — the pipelines, platforms and tooling around the models, not just the models themselves. Right now that means owning the LLM evaluation platform used by every AI team at a global consulting client.
Open to new roles, and happy to relocate
$ whoami
01 / The path
St Malachy's College, Belfast
Sat GCSE ICT three years early and got an A — the interest was settled well before anyone called it a career. Left with top of the school in Computer Science, AAB at A-Level, and All-Ireland medals from seven years of cross country.
PwC · 20 places from around 550 applicants
Four years of client delivery running concurrently with the degree. ASDA, Network Rail, Heineken — and volunteer work with a charity on the side.
Queen's University Belfast
PwC academic scholarship. Final-year project was the Stroke Rehabilitation Platform further down this page — four cooperating services, message bus, Kubernetes, delivered and graded.
Pytilia · three years contracted out to a US fintech company
Embedded at the client throughout, working in Python. Started on backend services, then built the notification system's entire E2E regression suite from scratch — around 300 tests running nightly. Ended up owning evaluation for the company's AI teams — the next two entries.
Ulster University · part-time, final year
Enrolled as a one-year PG Cert and extended it to a full MSc. Modules have covered machine learning, deep learning, robotics, knowledge engineering and MLOps on Azure.
Agentic AI · evals, MCP servers, platform
Ran a competitive vendor POC, made the build-vs-buy call, then stood up and ran the on-prem platform it landed on. Led a team of four.
Software engineering across the AI stack — platforms, pipelines, evaluation
Drawn to genuinely hard technical problems — the plan is to keep expanding the skill set for a long time yet.
02 / The work
A pluggable agent ladder behind one language-agnostic protocol: random, minimax, MCTS, and reference implementations in more than one language to prove the protocol actually holds.
In 2026 everybody has an LLM wrapper. Almost nobody hand-rolls tree search.
View on GitHubEnd-to-end training pipeline on Azure ML with CI/CD, versioned model deployments and a queryable endpoint. The side project closest to the day job.
View on GitHubGame-based physiotherapy assessment tracking hand-eye coordination over time: Expo front-end, .NET CQRS API, RabbitMQ worker services, Docker and Kubernetes packaging, SpecFlow BDD tests.
Final-year team project — and a complete one: four cooperating services, message bus and k8s manifests, delivered and graded.
View on GitHubROS 2 worker node navigating to a destination while avoiding obstacles, with the pathfinding written by hand and the whole stack packaged in Docker.
View on GitHubOffline-first PWA for learning Irish, with LLM-generated cards and real sync-conflict resolution. Genuinely unfinished, and v2 is in progress.
View on GitHubMore at github.com/thomas-c-reid
03 / Colophon
Three files. No framework, no build step, no dependencies — one HTML file, one stylesheet and about ninety lines of JavaScript. The moving gradient behind you is CSS, not WebGL.
It was built with AI assistance, and I'd rather say so than pretend otherwise — I evaluate these systems for a living. The interesting part wasn't the generation; it was the review. Every line that shipped got read, argued with, and cut where it deserved it.
It started out served from a Proxmox VM on the machine in my house, with a Cloudflare Tunnel carrying inbound requests so the router never faced the internet directly. That setup was one broadband outage away from being someone else's problem, so it now defaults to Vercel — a git push is the entire pipeline.
04 / Say hello
Open to roles in AI evaluation, LLMOps and agent engineering. Based in Belfast, happy to relocate.