I build and evaluate AI systems that have to survive review — agent evaluation harnesses, claim-support audits, and provenance pipelines.

The discipline comes from more than eight years in sterile pharma QA/QC (cGMP, USP 797): preserve the source state, make failure modes testable, and leave enough evidence for someone else to challenge the result. That background is my credibility, not the only audience I write for.

Based in Ontario, Canada. Bilingual English and French.

1. Claim Audit Lab — audits whether draft claims are supported by supplied evidence (deterministic rules, Markdown + JSON reports, explicit limits).
2. Evidence Bundler — nominates candidate passages into a reviewable bundle. Retrieval nominates; review decides.
3. Apparatus Contracts — versioned handoffs and integrity checks between those stages.

1. agent-eval-notes — public methods tour: harnesses that score better and are not promoted, multi-path coding screens, task-family transfer, RAG routes, verify-tool honesty.
2. verified-done — runnable demo: verified pass vs false completion vs scope violation, with a selftest and scrubbed live evidence.

Together, Path B treats scope and false "done" as first-class failures, not soft notes.

• On a sealed coding-agent suite, a packet harness improved verified passes 28/36 → 34/36 and cleared scope violations 6 → 0, then was not promoted because a pre-registered gate caught 2 false completion claims on multi-file work (exploratory; details on the tour).
• On a public demo split, holding the model fixed and removing a run_verify tool flipped honesty outcomes (and a local open-weight coder reproduced the flip; another model already abstained either way). Method: Report 06 + verified-done.

Numbers are exploratory measurements with stated n, not production validation.

• MainFrame — public Stage 1b cut of a lifecycle-first Markdown knowledge OS (inbox → ingest → knowledge → live → projects → archive). Nested MindGraph local hybrid retrieval (lexical + semantic + graph). Private corpora stay private.
• MindGraph — standalone package of the same engine when you only want the retriever.

• Career Decision Engine — browser decision-support tool with visible scoring and confidence labels. Live demo.
• Research Scaffold Harness / Basic Research Harness — scaffold and agent-loop experiments with provenance-oriented outputs.

These tools are not a truth engine, not a regulated quality system, and not a claim that production agents are “validated.” Evidence Bundler nominates. Claim Audit Lab audits relative to supplied evidence. Agent-eval numbers are sealed or demo-split measurements with stated limits. Software work is self-taught and project-based.

Work where AI systems meet review, reliability, and real operating constraints:

• AI evaluation, agent reliability, and evidence-handling workflows
• regulated software, quality systems, and data integrity
• forward-deployed or consulting work that maps a real process before automation

• LinkedIn
• Email