A selection of systems I have designed and put into production. Consulting follows the same shape: one workflow, measured, and owned by the team that has to run it.
I own the AI systems a company actually runs: retrieval over internal knowledge, document and metadata extraction, tool-using agents, and evaluation so a change can be judged against a baseline rather than a demo. Architecture, implementation, and ongoing ownership sit in the same role.
Retrieval over documents and structured sources, combining a graph database with a shared tool layer so context, search, and permissions were accessed the same way across services. Answers carried sources. Fallbacks covered the cases where retrieval missed.
Language-model services for metadata extraction, translation, and domain-specific reasoning, exposed through APIs on Kubernetes. Agents called search and other tools, with fallbacks when a step failed, and interfaces so non-engineering stakeholders could review outputs before they went further.
Event-driven pipelines for batch and low-latency workloads, from warehouse jobs through to streaming inference. Separately, an agent-based review of pull requests that flagged infrastructure conflicts against diffs, observability config, and tickets before they reached production.
When a language model is the wrong tool: medical imaging with an ensemble of vision models, object detection for industrial use, and end-to-end training and deployment pipelines a team can run after handover.
Bring one workflow. Thirty minutes is usually enough to see whether it is worth building.