Field notes on building and operating AI, automation, and infrastructure systems.

I’m Gavriel Sierra, based in Canada. I build these systems, operate them, and write up what happened.

How I approach this work

Before automating a workflow: a real workflow, a named owner, permitted data, a defined action boundary, and a way to decide whether the result should expand, change, or stop.

What I work on

Workflow and AI readiness

I examine where work is slow, repetitive, inconsistent, or difficult to review, then compare process, data, conventional automation, and AI options. The goal is one defensible next decision—not a technology shopping list.

Governed automation

I design tool and agent workflows with narrow capabilities, human approval for consequential actions, read-back evidence, and explicit failure and recovery paths.

Technical foundations

I build and operate Windows, macOS, and Linux environments, alongside Docker, model-gateway, observability, retrieval, secret-management, private-access, backup, and service-management patterns. I write about the decisions and tradeoffs without publishing the private infrastructure map.

Evidence and communication

I turn complex technical and career material into reviewable profiles, runbooks, decision records, dashboards, articles, and handoff artifacts. A polished result should make uncertainty clearer, not hide it.

Featured work

Featured writing

Why the smallest useful next move is a better place to start than an AI transformation program

“Transformation” may be a valid direction, but it is usually too vague to fund as the first unit of work. This article provides a practical framework for one useful, testable decision with an owner, data and action boundaries, acceptance tests, and an explicit expand, revise, hold, or stop decision.

Read the article · See all 17 articles

Working principles

Evidence before claims

Built, tested, live, historical, and proposed are different states. I try to say which one applies.

Human ownership at consequential boundaries

Models may help analyze, draft, and recommend. A person remains responsible for external communication, material changes, commercial commitments, risk acceptance, and other consequential decisions.

Recovery and handoff are part of the build

A system is easier to trust when someone can operate it, detect failure, restore it, export the data, and remove it without relying on the original builder’s memory.

Small enough to learn; complete enough to matter

A bounded first move still needs an owner, representative tests, an operating path, and a stop decision. “Small” should not mean disposable pilot theatre.