Brendan O'Brien, PhD

AI

Two separate things, and I do both: building language models into products that people rely on, and rebuilding how engineering itself works around them. I own my company's Claude deployment, and I ship production software as a solo engineer at a throughput that would not otherwise be available to me.

AI in production systems

Structured extraction — ReportHub

I designed and run the extraction pipeline for ReportHub, a clinical reporting platform in production with pathology laboratory customers: Claude via Amazon Bedrock pulling structured data out of pathology reports that arrive as PDFs, with output validation and prompt evaluation held to a production standard and measured against reference reports.

The interesting problem is not the extraction. It is knowing when the extraction is wrong, in a setting where a silent error propagates into a clinical document.

Conversational product features — Pluck

Pluck is a meal-planning app built on Expo, React Native and Supabase, with conversational recipe creation built directly on the Anthropic API. I took it from schema design through to TestFlight and Google Play beta.

Evaluation — PathReporter

For a speech-to-text dictation product used by pathologists, I built the production evaluation system: defining the metrics, assembling the reference sets, and monitoring transcription performance against them in production. Evaluation design is the part of this work I am strongest at, and it comes directly from twenty years of experimental science.

AI in the engineering process

claude-dev-loop

An autonomous multi-agent development pipeline I built. A coding agent develops features under enforced test-driven development, then hands off to specialised review agents — code quality, performance, security, infrastructure, compliance, test coverage — and iterates until every reviewer approves before a pull request is raised.

The constraint that makes it work is that the loop cannot exit by declaring itself finished. It exits when independent reviewers agree, which is a different and much harder bar.

Company-wide Claude deployment

I own the Claude rollout across Labflow: onboarding engineers and non-engineers, monitoring spend, and building usage dashboards across business units. That means I have seen where these tools deliver and where they quietly do not, across a whole organisation rather than one team.

Shipped with it

The clearest evidence is what exists. These are real products, in app stores and in production, built and maintained by one person alongside a full-time role.

I am not interested in the claim that AI writes the software. It does not. What it changes is how much surface area one engineer can hold in their head at once, and how cheap it becomes to build the thing you would otherwise have argued about. The judgement about what to build, what to measure, and whether the result is good enough remains entirely the engineer's, and that is the part I have spent a career on.