Work

A private-bank onboarding agent that streams the whole interface, not just text. Generative UI cards, risk gauges, and human approval desks over one SSE stream, orchestrated by a process engine. Runs without API keys in shadow mode.

One of the first working implementations of Google's AP2 v0.2 outside their own samples, wired to an external governance layer. A shopping agent buys FIFA tickets overnight under a cryptographically signed mandate it cannot override. Repo goes public when EP 03 ships.

Multi-agent mortgage pre-qualification with a judge agent gating decisions at a confidence threshold, and a locally-hosted research agent behind a data-privacy boundary. Became the worked example in Camunda's official A2A blog series and a Gemini 3 hackathon listing.

Built when nobody was asking for agent evals, now it is the first question in every room. Measures cost per case, accuracy against a gold set, tool-skip rates, and temperature sensitivity for process-embedded agents. Field-validated on live enterprise workloads across 20+ models. The method is public; the code is not.

A marketplace-submitted connector bringing Cohere's models (chat, structured JSON, classification, embeddings, reranking) into process orchestration as element templates. Built with sovereign-data deployments in mind, where a Canadian model and a self-managed cluster matter.

A deterministic linter that scores MCP tool specs 0 to 100 for agent-readiness and auto-fixes the failures. Born from eval findings that agents skip tools because of bad specs, not bad models.

The anatomy of the agentic loop, the primitive every major coding tool independently converged on, published as a spec, a deep-dive, and a motion graphic built with the loop it describes.