What I have built

A cloud-designed NVIDIA vision stack reworked to run on Jetson hardware on-premises, so GDPR-sensitive video is processed in the store and only analytics reach AWS. Became a reusable internal offering.

Argo CD for reproducible deployment, per-workload cost attribution with OpenCost, centralised logging and load testing, plus an Apache Iceberg data lake that avoided costly managed databases.

Development and production environments built from scratch for a chatbot on a live commercial portal, with CI/CD, a reusable Terraform template and keyless OIDC deployment into AWS.

An audit showed roughly 70% of a 50 to 60 second response time came from engineering rather than the model. Rebuilding the deployment architecture halved it at equal task performance, with a fine-tuned small model served on the client's own GPUs.

A self-hosted platform where effective access is the intersection of user and agent rights across C1-C4 data, with revocable delegation and auditable decisions deployed under GitOps on Kubernetes.

A Go CLI comparing inference serving configurations on latency, throughput and quality, internal GPUs operationalised behind Teleport, two AWS AgentCore production deployments, and mentoring in the open-source squad.

A departmental suite for a private clinic, designed to run on twelve Windows workstations with no route to the internet. PostgreSQL, Cerbos, and Zitadel replace the cloud data, authorization, and identity paths, while Google authentication remains an optional fallback.

A commercial agent and forecasting pipeline in telecommunications, a clustering pipeline over a three million member portfolio in insurance, a recommender in production in pharmaceuticals, and cloud deployment tooling built alongside the client work.

First engineering role, held alongside the machine learning engineering degree attached to it. Halved the mismatch rate of an entity recognition model and put it on scheduled retraining.