I build reliable AI systems that turn messy real-world workflows into tested, observable software — with explicit autonomy boundaries, evals, verification gates and human authority where it matters.
How I build: coding agents do much of the implementation. I design the problem, system architecture, specifications, autonomy boundaries, evals and evidence that make the result trustworthy.
Four projects that make the same engineering judgement visible: automate useful work, but keep permissions, uncertainty, evidence and consequential decisions inspectable.
Evidence-first control infrastructure for sensitive AI workflows: connect identity, policy and evidence, run deterministic checks outside the LLM, and preserve human authority before consequential actions.
Evidence-first identity resolution that returns merge / separate / review instead of forcing ambiguous organisations into a yes/no answer.
Turns deterministic synthetic administrative data into an operational monitoring brief across data quality, drift, statistical uncertainty, explanation and human review.
Reusable bounded AI workflows with capability gates, required evidence, human release and fail-closed behaviour instead of unrestricted tool execution.
Built and ran an Amazon business end to end across sourcing, logistics, listings, fulfilment and customer operations.
Coordinated 8+ staff and high-volume daily service, resolving operational bottlenecks under real-time pressure.
Worked across marketplaces, website, B2B partners and acquisition channels, using customer feedback to improve positioning and priorities.
I design agentic software-development systems with explicit specifications, autonomy boundaries, evals, verification gates and production monitoring.
The portfolio is built to let a reviewer verify that claim: working interactions first, then evidence, code, tests and explicit boundaries for what is synthetic, measured, human-reviewed or not yet proven.