I build LLM systems for regulated sectors, and the evidence layer that lets others audit them: evaluation harnesses, audit trails, AI Act transparency, cost controls.
LV / PRACTICE FILE 2026Utrecht, Netherlands / Seven years in EU-regulated sectors
answer_id a4f1c2e0-9b73-4d1a-8e55-71c0f2ab9d34model_version pinned, recorded per answersystem_prompt v14 (live), author != approvertools_called 3 of 26 permitted, schema-validatedfree_form_sql none permittedtenant_scope re-verified server-sidetokens prompt 4,812 / completion 611All figures are from work delivered first-hand.
On recordMigrating 246 pipelines and 67 workflows off legacy ETL tooling
Across 44 tenant schemas, counting player-level data alongside the analytics facts
Down from roughly three hours after re-platforming onto 43 parallel tasks. The player-data pipeline came down from nine hours to three.
Production commits from AI-assisted development, each logged with pinned model versions and credential reads blocked at the tool level
AWS Foundational Security Best Practices and CIS benchmark conformance on the infrastructure I operate; independent penetration test, findings remediated and retested
Evaluation harnesses, audit trails, Article 50 transparency, agent guardrails and disclosure controls. The evidence a regulator or an enterprise buyer actually asks for, generated by the system itself.
Retrieval, structured output through forced tool-use rather than parsed prose, regression evaluation against versioned baselines, per-call cost accounting, multi-provider failover.
Multi-tenant warehouse pipelines and migrations, RBAC and row-level security, append-only audit logging. The substrate every control above depends on.
20-case
eval harness
Vector retrieval, structured output through forced tool-use, tagged evaluation baselines with a regression-diff tool, per-call cost accounting persisted per decision, and multi-provider failover.
+5,015
lines, one release
A typed permission-scope vocabulary, page and action level role guards, an append-only audit log enforced by database trigger, and Postgres row-level-security tenant isolation.
246
pipelines migrated
Sole-authored migration off legacy tooling across 453 commits, then re-platformed onto 43 parallel containerised tasks: roughly three hours down to one hour forty.
Art. 50
in shipped code
EU AI Act transparency reasoning at the point of interaction, guardrails blocking agents from reading credentials, least-privilege tool allow-lists, and backups taken outside model control.
This is the shape of a single record from a governed LLM system: the prompt version, the model that produced it, the tools it was allowed to call, what it cost, and who signed off on the rule that constrained it. Not a policy describing the control. The control, emitting its own proof.
The governed answer path. Flagged nodes are the ones a regulator asks about first.
Random-forest classifiers predicting product potential with experiment tracking, unsupervised outlier detection flagging irregular terminal performance, churn and high-value customer models, and market-basket analysis using association rules and Markov transition matrices.
An image-generation service producing packshots for an online store, with a synchronous and batch path and memory ceilings so it survives constrained hosting.
Product and price extraction from Instagram captions and order threads, running under a strict contract: totals are quoted verbatim from the source, identity fields resolve server-side, and anything uncertain routes to a human rather than to the model's best guess.
The model reads; it does not decide.
Engagements start with a scoping call. You work with the person who does the work.