I turn operational requirements into working systems. My background spans country-office and regional operations, finance processes, procurement and EU project delivery. I now bring that experience into AI-assisted application development.
I own the problem definition, architecture and implementation direction. I work with Claude Code, Codex and VS Code to build, inspect and test the software. Human-directed, AI-assisted engineering is part of the method, not something I hide.
Evidence-linked workflows, bounded context, specialist analysis and model-provider interfaces. I care about what happens when a source is missing, a model fails or an action needs human authority. A generated answer is not the same as a completed process.
My architectural stance is practical. The model should be replaceable, existing tools should be reused where they fit, and consequential actions need explicit boundaries. Tests and audit records should make those boundaries inspectable.
Local-first reference shell demonstrates a bring-your-own-model interface, a human verification queue, allow-listed actions and explicit no-provider behaviour. It includes runnable verification and synthetic demo data. It is a reference implementation, not a claim of a production multi-agent execution service.
Architecture and reproducible simulations publishes the design, assumptions, methods and limits. The assessment and simulations were produced with Claude under my direction. They are not an independent third-party audit, and simulated results are not realised business ROI.
I distinguish source implementation, fresh tests, historical execution and actual user outcomes. Token volume, capability inventories and passing simulations do not establish useful adoption or commercial impact.
The work I want to do next is business-owned AI operations. Understand the process, build alongside its owner, measure the complete result and leave the team able to run it.
On 9 October 2026, the initial clean-checkout review of the public reference shell passed 160 tests and eight separate verification checks. A later review that day hardened its risk-score input boundary. The current suite passes 384 tests: the original 160 plus 224 parametrized input-domain, valid-boundary and CLI regressions. The eight-check verifier also passes. Provider fixtures are mocked, and the HTTP smoke check uses isolated synthetic data. These results establish their stated local contracts, not live model quality, production security, user adoption or measured business ROI.
The current public CI run, tied to the published implementation commit, completed successfully on 9 October 2026 for Python tests, verification and the UI build with its production dependency audit. A passing dependency audit is a dated package check, not certification of the whole application.