Can your data carry agents? Lineage, entity resolution, governance and quality, judged against what agents actually need. Thirty years of data and architecture behind one question.
See the proof →The credibility layer that separates a consultancy from a CV, and the receipts for learning by doing: every project exists to find what doesn't work as advertised.
Foreman keeps AI-written software honest. A feature passes three gates only I sign, and an agent does the work in a sealed room: a hard cap on attempts, escalate instead of guess. The process is public. The judgment stays private, because that is the part nobody can copy.
Flow · Sprint · Expertise PacksA sustainability benchmarking engine on public emission registers: 8 national PRTRs, 4.8 million rows, row-level lineage, running in Azure. The hard part is not the pipeline, it is entity resolution: plants are named after sites, companies after legal entities, and nothing joins for free. The data has a hard ceiling, and that finding is worth more than the product would have been.
Azure SQL · Python · Public registersThe small tools that keep a one-person company honest. BurnRate measures real token spend from session logs: 95-98 percent of input was cached context, not fresh work. Voice-driven Time Registration turns spoken sessions into an audit-grade hour log. SentinelPR and ReleaseScribe, two open-source GitHub Actions, round out the drawer.
Python · JSON stores · SQLCipherFor teams deciding what to do about AI who want a tested, independent opinion instead of a vendor pitch. You bring the decision. I bring what broke when I built the equivalent system myself.
Half or full day, on site or remote, plans and architecture checked against what has actually been built and measured, ends in a short written opinion.
45-60 minutes for engineering audiences, built from real experiments, not a vendor deck.
A small PoC that answers the one risky question, with numbers, then ends.
An independent read on a vendor, an acquisition target, or a build proposal, before budget commits to it.
No implementation contracts, no staffing, no reselling somebody else's software. A scoped PoC to settle a decision is fair game, a team on the ground for six months is not. I would rather say that now than after the first invoice.
Measured 85-98% of token spend in long agentic sessions was overhead, then cut it with concrete session-management fixes.
Built a multi-country emissions data pipeline: eight national registers plus the EU-wide E-PRTR, 4.8 million verified rows with row-level lineage.
Run the company on a portable file-based control plane; a voice agent on an 8 GB laptop GPU is one of its surfaces. Moved the whole tree across machines and AI subscriptions in an afternoon, nothing lost.
Cut design-review time by roughly 84% with a self-tested expertise agent, and threw out its perfect self-test score after auditing the test itself.
Let an agent take a data-pipeline change from spec to merged PR overnight, 73 tests green, while humans kept push, merge and declare-done.
Thirty years in technology. Took a Competitive Market Intelligence platform at Valona Intelligence from a Gmail inbox and a spreadsheet to a six-layer, multi-tenant SaaS product serving hundreds of enterprise customers, engineering team from 2 to 28 over eight years without accumulating legacy. That platform is now a Leader in the 2026 Gartner® Magic Quadrant™ for Competitive and Market Intelligence Platforms. At Damen Shipyards, turned a solo big-data and IoT initiative into a CIO-sponsored, 30-person program that became its own business unit, Damen Digital. Advised HuurPrijsHulp, an Amsterdam start-up named third most innovative social-impact company in the Netherlands at the 2024 KVK Innovatie Top 100.
Different project, different language, 17 days apart. The mechanism is boring on purpose: a failure written down as a checklist question outlives the fix, and the next reviewer, human or agent, has to answer it.
The bottleneck in working with agents was never the decision. It was being the messenger between two machines. Here is what happened when I removed the relay and kept the decision, including the moment my own approval got thrown out.
One person plus agents is a real operating model in 2026. The design question is not what the agent can do, but which verbs you refuse to delegate: push, merge, declare done.
Advisory enquiries, talk invitations, or questions about the tools. I read everything.
