Scott
Clark

There is a particular kind of intellectual restlessness that earns a PhD in Applied Mathematics from Cornell, three simultaneous undergraduate degrees from Oregon State, accumulates 1,200 academic citations before the age of 35 - and still walks away from a VP role at Intel to go back to zero. Scott Clark has it in spades.

His company, Distributional, is not easy to explain at a cocktail party. It is the platform that tells enterprise teams when their AI systems have quietly become something different from what they deployed. Not if a model produces wrong answers - any eval can catch that. But whether it has drifted, shifted, hallucinated in new patterns, or quietly broken in ways that no benchmark anticipated. Clark calls these the "unknown unknowns." The industry calls them expensive surprises.

Distributional's approach: behavioral fingerprinting. Every AI agent or model leaves a statistical signature across its outputs - a distribution of behaviors. When that distribution shifts, the system flags it. When it clusters differently, the system asks why. It is, essentially, bringing the rigor of the academic hypothesis test to production AI systems that have never seen one.

Clark has proposed a framework for this called Maslow's Hierarchy of Observability for AI - a deliberate provocation aimed at an industry still arguing about whether log files count as monitoring. Unsurprisingly, it landed.

The career arc reads cleanly in retrospect. At Yelp from 2012 to 2014, Clark built the Metric Optimization Engine (MOE) - a Bayesian optimization tool that would become the conceptual seedling for his next decade. He also launched the Yelp Dataset Challenge, seeding ML researchers with real-world data at a moment when real-world data was scarce and precious.

Then came SigOpt in 2014 - a startup purpose-built around Bayesian optimization for machine learning models. Y Combinator backed it early. Andreessen Horowitz followed. Seven years later, Intel acquired SigOpt in October 2020, and Clark stepped into a role most founders never see: VP & General Manager of AI and HPC Supercomputing, a budget of $70M, a team of 200 engineers, and all the institutional weight that comes with being inside one of the world's largest semiconductor companies.

He lasted three years. Not because it went badly, but because a cleaner problem presented itself. AI systems were being deployed at enterprise scale without any coherent framework for understanding what they were actually doing after launch. The testing tooling was primitive. The observability was often non-existent. And Clark - who had spent two decades thinking about optimization, statistics, and the behavior of complex systems - saw exactly what needed to be built.

Distributional launched in September 2023. By December of that year, Andreessen Horowitz had led an $11M seed round - yes, the same firm that had backed SigOpt, which is the kind of continuity that says something about a founder's track record. Less than a year later, in October 2024, Two Sigma Ventures led a $19M Series A, bringing the total to $30M.

The founding team is a masterclass in network effects from two startup cycles. Members arrived from Bloomberg, Google, Meta, Intel, SigOpt, Slack, Stripe, Uber, and Yelp. Collectively, 24 people carry 35 technical degrees and five PhDs. Clark, characteristically, describes the team structure with the precision of someone who once wrote a dissertation titled "Parallel Machine Learning Algorithms in Bioinformatics and Global Optimization."

The product is built for enterprise: deployable in VPC, locally, or on Kubernetes; integrated via OpenTelemetry, SQL, or Distributional's own SDK; with dashboards designed for multi-functional teams who want to understand AI behavior without needing a PhD to interpret it. Clark has managed the paradox with characteristic care - rigorous under the hood, accessible at the surface.

He appears on podcasts now, explains why AI agents fail in ways your evals don't predict, and maps production traces into vector fingerprints to uncover emergent behaviors. He says it simply. He means it precisely. There is no gap between those two things, which is rarer than it sounds.

In May 2025, Oregon State University spotlighted him as an alum who had built a $30M AI startup. Three undergrad degrees completed magna cum laude in four years. The timeline makes more sense when you understand that Clark does not treat constraints as obstacles. He treats them as specifications.