Skip to main content
matthewpurdon
Matthew Purdon

Principal Engineer & AI Platform Builder · Toronto, Canada

Matthew Purdon

I build the systems around AI-assisted engineering: platform boundaries, review loops, agent workflows, and the team habits that keep generated software from turning into entropy. And I lead the teams who run them.

Twenty-five years in software, writing Field Notes on the new SDLC, code review, hiring for judgement, and how teams actually ship.

01

The Work

My work sits at the intersection of architecture, process, and AI-assisted delivery. I care less about demos and more about the operating system that lets teams ship with models in the loop without lowering their standards.

Writing principal engineering, review, AI-assisted delivery
Building agentic tools, MCP servers, local-first review systems
Arguing review is now the main engineering act
Based Toronto, Canada
02

Start Here

Judgement is the job now

AI removed the effort filter. Taste, review, and technical judgement are what remain.

What is a principal engineer?

Principal engineering is not one archetype. It is the ability to shift between all four without losing the thread.

Interviewing the AI-assisted engineer

Syntax under pressure is a dead proxy. Hire for the AI era by testing judgement and taste: a real task, a few quiet traps, and the review round.
03

Recent Field Notes

Teams & Process · 9 min read

You cannot punish your way to a great culture

Every few months a leadership post makes the rounds insisting culture is what you tolerate. I keep landing on the opposite, from humility more than pride: the best teams I have watched ran on trust. On why punishment only buys you silence, the unglamorous work that actually grows a culture, and why it comes back to who you hired.

Scarcity was the feature

Every sovereign AI pitch in this country opens with the same apology: Canada cannot out-spend the hyperscalers. But scarcity built better engineers once, and it can again. On metered compute, capabilities over models, and a library card for the model era.

Postman was a workaround

I set out to explore a sparsely documented API in Postman and lasted about an hour before I gave up and built the client I actually wanted. The generic tool was only ever a workaround for a cost that just evaporated: building the specific thing yourself. On why buy-over-build is dying, why most of your code is now disposable, and the access-patterns-to-spec-to-code ritual I build with now.

Build the model a map

An agent will rediscover your whole database every session, and pay for it. The fix is not a better prompt; it is writing the map down. On indexes, stubs, skills, and the documentation your teammates never got.
04

Selected Lab Reports

SoftwareAI-assisted

Grey Eminence

A native macOS meeting recorder that never joins the call: on-device transcription, speaker diarization, and a Claude-powered memory of everything said.

View report →
SoftwareAI-assisted

TCC

An experiment with the tiny pi coding agent that became my daily-driver harness: AWS Bedrock underneath, twenty-five extensions on top, every behaviour mine to change.

View report →
SoftwareAI-assisted

MCP servers

A monorepo of Model Context Protocol servers built as facades, not adapters: each tool answers a question I actually ask, with the org's tribal knowledge baked in.

View report →
05

Browse Topics

#AI-Assisted Engineering#The New SDLC#Product#Hiring#Teams & Process#Industry#Sovereign AI#Opinion#Notes