Selected Work

A large organisation needed to reduce time spent on routine internal requests across departments. Manual handling was slow, inconsistent, and a poor use of staff capacity.

I built an end-to-end AI assistant that handles incoming requests, retrieves relevant information from internal knowledge bases, and routes complex cases appropriately. The system runs on cloud infrastructure I designed and deployed.

A professional services firm was spending significant budget on manual document review and data extraction. They needed to know if AI could replace this reliably — and if so, ship it.

Started with a small PoC to validate quality and ROI. Once confirmed, moved to a production-grade system with improved extraction accuracy and full CI/CD from day one.

AI features were adopted quickly, but model and retrieval costs could not be reliably attributed to individual product features or customers. That made product decisions harder and allowed cost increases to surface too late.

I built a FinOps foundation for GenAI workloads: consistent tagging, cost allocation, usage analysis, budgets, and alerts. This produced concrete optimisation work for model selection, retrieval, and tenant-level limits.

Several GenAI prototypes existed, but changes to prompts, models, and retrieval could not be assessed consistently. The team needed a practical way to make release decisions without turning every iteration into a manual review.

I designed an evaluation and governance layer around the existing AI applications: representative test datasets, automated quality checks, Guardrails, release gates, and production monitoring for quality, latency, and cost. The result is a repeatable path from experiment to controlled production release.

A technology company had already invested in AI features but wasn't confident the architecture would scale. They needed an outside assessment before committing further.

In a short, focused engagement I reviewed their current approach, identified the critical risks, and delivered clear prioritised recommendations — no long commitment required.