I work with the people who will use the system. I stay when it goes live.
01 / Practice
CONTEXT
I find useful business problems with stakeholders, shape the product, build it, and stay involved through acceptance.
Useful model output depends on context, tool boundaries, state, evidence, and clear interfaces.
The implementation covers retrieval, ranking, agent workflows, evaluation, observability, privacy, and production verification.
02 / Systems
I turn model pipelines into usable production systems.
Cortex connects a Next.js interface and FastAPI backend to recruitment data. The source records stay read-only. Recruiter feedback is stored separately, and each user sees only the records they own.
Role
Product + architecture + code
Boundary
Upstream records read-only
Span
CRM data to recruiter decision
03 / Record
Selected evidence
Production systemPublic / linked
Cortex
Rebuilt the recruiting system at Pauwels Consulting around AI. Recruiters get ranked candidates for every vacancy with the reasons shown, and a shortlist that took an afternoon takes minutes. I shaped it with the recruiters who use it, then sat with them through the switch, which took as long as building it. Every shortlist they accept or change feeds back into the ranking.
Launched Seld, a local-first, open-source AI chief of staff that keeps open work, decisions, evidence references, and next actions in plain files so agent sessions can be replaced without losing the outcome. I built the product architecture, deterministic continuity layer, public site, and evidence-led release process.
Released codex-multiprovider 0.1.0 alpha, which lets Codex use Kimi K3 and experimental Claude routes through an external CLI bridge. It includes fail-closed routing, a gated continuity preview for provider exhaustion, Fable review, and private ChatGPT Pro handoffs without changing the signed app.
Python / Codex CLI / Provider routing / Local continuity
Diagnosed a Claude Code 2.1.0 release crash, traced it to semver parsing of a dated changelog header, and submitted a minimal repair. Anthropic confirmed the fix and closed the report.