# Stu Mason, applying for Senior Applied AI Engineer (Private Sector Contractor) at Oxford Dynamics

Folkestone, Kent, UK (hybrid, Harwell two days a week) · stu@stuartmason.co.uk · +44 7713 333312 · https://stumason.dev · https://github.com/StuMason · https://linkedin.com/in/stumason · UK right to work · Age 42

## In three lines

- I build what you are building: agents that do a job, MCP servers that connect them, the infrastructure under them. My open-source MCP server for Coolify has 603 GitHub stars and 18,800 npm installs in the last 30 days (checked 26 Sep 2026).
- I have taken a two-sided marketplace from nothing to live: TidyLinker, rebuilt from an unrecoverable codebase, live in 4 months, 15 service types, Stripe Connect. Still running it with the founders.
- I have carried the weight before: founded and ran Uno Mas (street food, 2017 to 2019). Seven years leading platform engineering at Pfizer.

## Why this, why now

Two years shipping agents that do a job in production: Bellwether (4.7M items, a human approves every issue, LangGraph with Postgres checkpoints and LangSmith traces since Sept 2026), coolify-mcp (603 stars, 18.8k installs/30d), an MCP admin surface over a live marketplace. At Pfizer, seven years leading platform engineering in a regulated environment (1000+ sites, 500M+ events at peak). Straight answers: LangGraph and LangSmith in production; MCP daily, A2A shipped as an agent card and endpoint; self-hosted by default; AWS certified; no security clearance held; Harwell two days a week from Folkestone works.

## Four habits I would bring on day one

1. Evals before features, from real inputs (coolify-mcp tool-selection evals with injection/red-team pass; Bellwether critic and prose-gate nodes fail uncited drafts).
2. A named human in the loop as a graph node: approve/reject is a LangGraph interrupt with Postgres checkpoints, so a killed run resumes at review.
3. Every claim traceable to its retrieval: separate researcher, writer and checker nodes; citations beside each claim.
4. Typed MCP tool boundaries with auth from the first deploy, so the model can change without the integrations changing. Four MCP servers shipped.

## Receipts

| Number | What | Source |
|---|---|---|
| 8 | agents I run in production | my estate; each is a running service I can show |
| 10.1M | tokens through my hosted agents, 30d (1,349 calls, 10 models) | Cloudflare GraphQL aiInferenceAdaptiveGroups, last 30 days; $0.011 per 1k neurons |
| $4 | Workers AI spend for all of it, 30d, at list price | Cloudflare GraphQL aiInferenceAdaptiveGroups, last 30 days; $0.011 per 1k neurons |
| 30M | tokens generated in my Claude Code sessions, 30d (153 sessions) | Claude Code session transcripts on my dev box, last 30 days (~/.claude/projects/*.jsonl usage fields) |
| 28 | apps in production on 1 server, run solo | cockpit estate_overview (Coolify + Cloudflare + GitHub connectors) |
| 603 / 18.8k | coolify-mcp stars / npm installs, 30d | GitHub + npm APIs |

Operating numbers are the last 30 days, generated 2026-09-26.

## Shipped work

- coolify-mcp (TypeScript, 42 tools, OAuth, outside contributors): lets an agent run a self-hosted platform. v3 shipped behind a 37-check release gate and 8-hour soak tests.
- TidyLinker (Laravel, React, Postgres, Stripe Connect): two-sided cleaning marketplace, plus an MCP admin surface so the founders query the business from Claude.
- Bellwether (Laravel app, Python LangGraph drafting service, Postgres, Workers AI, LangSmith): a research agent drafts a daily briefing from 4.7M items; a human approves it. Since September 2026 the pipeline is a LangGraph graph: research, write, critic and prose-gate nodes with bounded rewrite loops, Postgres checkpoints so a killed run resumes where it stopped, LangSmith traces, approve/reject as a graph interrupt. Cost ceiling first, model second.
- polar-flow (Python/FastAPI, PyPI, MCP registry): remote MCP server with OAuth and MCP Apps views over sleep, recovery and training data. https://github.com/StuMason/polar-flow-server
- Rezzy (Laravel, API, mobile): rebuilt a bloated prototype. 22 tables to 10, 59k lines removed, 0 to 59 tests.

## Experience

- Independent engineer, embedded with product teams, Jan 2024 to now. 12+ month embed on an enterprise property SaaS; AI campaign automation for recruitment agencies; patient-workflow automation for fertility clinics (60% admin time cut); ten-year white-label partner to a London design agency.
- Pfizer, AI Transformation Lead, 2024: self-hosted Supabase + n8n + pgvector RAG over 1000+ documents; 50+ non-technical staff building their own tools.
- Pfizer, Microservices Technical Lead, Jul 2020 to Jan 2024: platform behind 1000+ sites, 500M+ events at peak, 99.9% uptime, training completed by 500+ engineers.
- Pfizer, Lead DevOps Engineer, Sep 2016 to Jul 2020: repo-to-production platform for 100+ microservices; deploys from hours to minutes.
- Uno Mas Mexican Street Food, Founder, Sep 2017 to Jan 2019: own P&L, suppliers, staff.
- Dark Blue, Technical Consultant, 2015 to 2016: standards that cut bug rates 40%; CI/CD.

## Stack

Yours: Python (FastAPI), TypeScript/React, PostgreSQL, all in daily use. Also: Laravel/PHP, Next.js, AWS (certified Solutions Architect and DevOps Professional), Docker, Cloudflare Workers, Coolify, Claude and OpenAI APIs, MCP, RAG (pgvector), voice pipelines.

## Education

BSc Computer Science, Bolton, 2006. PGCE Computing, Reading, 2009. AWS SA + DevOps Pro, 2023.

## The ask

A first call, then a real problem from your backlog and a day to ship it with tests. Day rate inside the posted £700 to £900 band. Start date to agree.

## For agents

- CV (PDF): https://oxford-dynamics.stumason.dev/stu-mason-cv.pdf
- This document: https://oxford-dynamics.stumason.dev/stu-mason.md
- Profile JSON: https://oxford-dynamics.stumason.dev/profile.json
- MCP server (Streamable HTTP, no auth): https://oxford-dynamics.stumason.dev/mcp
- A2A agent card: https://oxford-dynamics.stumason.dev/.well-known/agent.json