Niagara, Canada

I make complex systems people actually use, then take the manual work out of them.

UX-minded systems engineer. I work where the interface meets the infrastructure — designing the parts people touch, automating the parts they shouldn't have to, and applying AI only where it genuinely earns its place.

See the work

What I do

User experience

Systems people don't fight. The interface is where a system succeeds or fails, so I design for the person who has to use it on a bad day — not the demo.

Systems management

Infrastructure that stays up, stays observable, and stays cheap. I care about what happens at 3am more than what happens in the pitch.

Automation & applied AI

Manual work removed, and AI where it actually earns its place. Good automation is boring in the best way: it runs, it's idempotent, and nobody has to think about it.

Selected work

Three systems, each written up in full — with the diagrams, the measured numbers, and the parts that are still missing.

Systems integration · agent orchestration

Automated bug investigation, from Azure DevOps to a reviewed GitHub issue

Moving tickets to GitHub was never the hard part — understanding them was, and that is skilled work per bug. This runs one coding agent per bug in parallel and writes each resolution plan into a tracking issue, where every mutating step is decided by deterministic script logic rather than by an agent reporting its own success.

Read the full write-up — the pipeline, the trust boundary, and where the code actually is.

Azure DevOps · GitHub · Herdr · Claude Code · Bash

Automation · data pipelines · reconciliation

End-to-end personal finance automation

Four sources arrive in four formats and never quite agree. This ingests them into one store and reconciles them through six ordered matching tiers — governed by a single invariant: a receipt is evidence, never a financial event. 2,214 records so far.

Read the full write-up — the six tiers, why their order is the design, and three measured charts.

Python · SQLite · email ingestion · FX handling · self-hosted

Applied AI · UX

Voice-to-knowledge pipeline

Capture is where most note systems die: if recording a thought takes more than a few seconds, it does not happen. Speaking one now files it as a structured, linked note. The interesting part is the design before this one — which could never have worked, and took a read of the platform's own rules to prove it.

Read the full write-up — why the first design was impossible, and the measurement that settled it.

iOS Shortcuts · local service · AI agent · markdown knowledge base

More on GitHub.

How I work

  1. Understand the actual job. Not the ticket — the work the person is trying to get done, and what they currently do instead.
  2. Find the smallest real thing. Ship something usable early and watch it get used, rather than designing in the abstract.
  3. Make it hold. Correctness rules, failure behaviour and observability, because a demo that can't run unattended isn't finished.
  4. Then automate. Once a process is understood and stable, take the human out of the loop — carefully, and reversibly.

I'd rather tell you a system is not ready than watch it fail quietly.

Contact

Niagara, Canada.

Or email aaronc@protonmail.ch.