The team
We’ve spent the last two years inside one of the largest enterprise agent rollouts, building the evaluation infrastructure that decides whether Microsoft Copilot products are safe to ship and watching where existing systems break. Novarch turns that discipline into a runtime layer your customers can put in front of their own agents: pinned models, cited rules, structured evidence, and an operator in the loop on borderline cases.
Co-founder
CEO
Spent three years as a PM at Microsoft Power BI, where he led the evaluation program for its Copilot and shipped AI features to millions of users. Featured speaker at FabCon 2024, 2025, and 2026.
Peer-reviewed ML researcher with an MS in Machine Learning: first author on a CogSci 2024 paper measuring how well deep-learning models capture human concepts, run across 18 model architectures.
LinkedIn
Co-founder
CTO
On Microsoft’s Security AI Research team, she builds agent-evaluation infrastructure for the enterprise. The LLM evaluation infrastructure she built there became part of Azure AI Foundry. Production agent-eval infrastructure for an enterprise audience is something very few people in the industry have shipped.
Second author on CTI-REALM, a published agentic security benchmark (March 2026). Microsoft’s EVP of Security cited it publicly when announcing the Project Glasswing collaboration with Anthropic.
LinkedInAdvisors
Former SVP & Chief Data Scientist, Equifax
At Equifax he ran global AI/ML research, the data science labs, IP development, and the work on explainable AI. In credit decisioning, a model that cannot say why it decided something is one you cannot put in front of a regulator.
Three decades across high-growth startups and Fortune 100 enterprises: productizing emerging technology, scaling R&D and engineering teams, and shipping under regulatory constraint.
LinkedInWhy this team
The load-bearing decisions in Novarch all came from the same place: years spent asking whether an AI system is good enough to ship into a workflow that matters, and seeing evaluation used to react to failures after the fact instead of stopping them before they ship. That’s where one call per action, a pinned model SHA, structured output with a cited rule and signals, a database-rendered audit document, and an operator on borderline cases all come from.
Runtime enforcement is downstream of that question. Novarch is the version of the answer your customers can run on agents you didn’t train.