Hi, I'm Padraig.
I build AI systems that survive contact with real users.
An AI engineer with three years owning production LLM, RAG and computer-vision systems end to end for enterprise customers. Now doing an MSc in AI at Edinburgh, and open to roles in London, the US and beyond.
A bit about me
I'm an AI engineer doing an MSc in Artificial Intelligence at the University of Edinburgh. Before that I spent three years owning production machine learning at yieldHUB: computer vision, anomaly detection and LLM agents running against live manufacturing data for enterprise customers.
I care about the unglamorous half of the job: the failure modes, the human-in-the-loop overrides, the evaluation that tells you whether you actually made things better. I write about it on Substack, and I'm always building something, most recently a local-first running coach on top of Strava.
Google & Kaggle, AI Agents Intensive Course
What I've actually moved
Numbers from systems I shipped against live manufacturing data and real customers, not benchmarks.
80%
reduction in manual review
X-ray microchip inspection CV system
16×
data ingestion throughput
pipeline redesign under tight constraints
60%
compute cost reduction
same pipeline redesign
Daily
anomaly detection at manufacturing scale
first production ML capability, automotive quality standards
Where I've shipped
Data Scientist (AI/ML Systems Engineering) · yieldHUB
Sept 2023 – Aug 2026 · Limerick
- Designed and delivered production AI (computer vision, anomaly detection, LLM workflows) inside a commercial analytics platform.
- Led a modular MCP-based agent architecture, reusable LLM abstractions aligned to existing APIs.
- Owned an end-to-end production X-ray microchip inspection CV system → ~80% less manual review.
- Delivered the company's first production ML (anomaly detection over large-scale manufacturing data, daily; automotive quality standards).
- Redesigned the ingestion pipeline → 16× throughput, ~60% lower compute cost.
- Built human-in-the-loop overrides and explicit failure-mode handling into production systems.
- Worked with customers in weekly R&D sessions, presented to international enterprise stakeholders, and mentored interns.
Data Engineering Intern · Groupon
Mar 2022 – Aug 2022 · Remote
- Built a GUI tool automating GCP data-transfer workflows for non-technical users.
- Built an Airflow-scheduled automated backup system.
Education
MSc Artificial Intelligence · University of Edinburgh
Sept 2026 – Present · Edinburgh
BSc Data Science & AI (2:1) · University College Cork
Sept 2019 – May 2023 · Cork
Python / SQL / R; machine learning (ensembles, feature engineering, tuning, disciplined validation). Statistical inference, hypothesis testing, probabilistic modelling, dimensionality reduction.
Things I've built
Three I've written up in depth, plus the labs I keep building to pressure-test how models fail.
RAG Eval & Observability
A full-stack platform for building and evaluating RAG systems, with the eval and observability rails built in.
- An ingestion pipeline that chunks, embeds and indexes documents into PostgreSQL/pgvector.
- Multiple retrieval strategies behind one interface, so they can be swapped and compared.
- A chat UI for interactive use, backed by a FastAPI service.
- An offline evaluation harness computing Hit@K and MRR over labelled sets, plus LLM-judge scoring for answer quality.
LLM planning over deterministic analytics, with full traceability and evaluation on top of MCP tool execution.
A multi-agent system that turns raw Git history into audience-specific changelogs and reports.
Also building
Model Failure Lab
A CLI that diffs two versions of an LLM/RAG system, shows what regressed, and turns each failure into a permanent test.
Agent QA Mentor
Agentic QA over tool-using agent trajectories, with 5-dimension scoring and a prompt-improvement memory.
Agentic Job Search Platform
Resume parsing plus Qdrant semantic search behind a multi-LLM FastAPI backend.
NanoGPT From Scratch
A from-scratch GPT-2 124M reproduction in PyTorch, with ablations and an interactive results site.
Most of the work is what happens when the model is wrong
Demos pass on the happy path. Production doesn't get that luxury. Four things I hold to.
Robustness over demos
I build for the messy middle, ambiguous inputs, partial data, edge cases. A system that only works on the happy path isn't finished.
Explicit failure handling
Every model and agent has failure modes. I make them visible, contain them with guardrails and heuristic fallbacks, and degrade gracefully instead of silently.
Human-in-the-loop
Production AI earns trust by keeping people in control. I design override paths and review surfaces so humans stay the final authority where it matters.
Evaluation as infrastructure
If you can't measure it, you can't ship it safely. Offline metrics, LLM-judge scoring and live observability are part of the build, not an afterthought.
The full stack of how I build
From agent architectures down to the pipelines and reliability work that keep them standing.
AI systems & agents
Vision & modelling
Data & retrieval / infra
Production reliability
What I'm writing & shipping
Notes on building AI systems, and how much I'm pushing week to week.
1,557
contributions in the last year
@Padraigobrien08Google & Kaggle, AI Agents Intensive Course
credential
I'm always up for a hard problem
If you're building something that has to work in production, I'd like to hear about it. I'm in Edinburgh for my MSc, open to London, the US and elsewhere, and email reaches me fastest.