Jason
Lopatecki

The Debugging Problem at the Heart of the AI Boom

There is a specific frustration that hits every engineering team around six months into deploying an AI product. The model works fine in testing. The demos are clean. The investor updates are optimistic. Then something goes wrong in production, and nobody can explain why. The outputs look plausible but wrong. Confidence scores say one thing; users say another. The system fails quietly, at 3am, in a way that only shows up in a support ticket two weeks later.

Jason Lopatecki built a company around that frustration. Not because it was a clever market thesis, but because he had lived it. At TubeMogul, the programmatic advertising platform he co-founded and took public before Adobe acquired it for over $500 million, he watched data pipelines and algorithmic systems behave in ways that were genuinely hard to understand in real time. The tooling to observe, explain, and repair those systems was always lagging behind the systems themselves.

When generative AI arrived at scale, Lopatecki recognized the same pattern - only bigger, faster, and with much higher stakes. He co-founded Arize AI in 2020 alongside Aparna Dhinakaran to build the observability infrastructure that modern AI systems fundamentally lack.

A Berkeley Kid Who Built Chips First

Lopatecki was admitted to UC Berkeley while still in high school. He graduated with a degree in Electrical Engineering and Computer Science - the kind of education where you learn both how systems are built at the hardware level and how to think about software at scale. He started his career at Alcatel and Calix doing chip design and systems architecture, which gave him an unusually grounded view of how software actually runs. Not everyone who runs an AI company has debugged silicon.

His first founding moment came with Illumenix, a video analytics company he built before TubeMogul absorbed it in 2008. At TubeMogul, serving as Chief Strategy and Innovation Officer, he helped pioneer programmatic video advertising - a move that restructured how brands spent billions of dollars on digital video. The company listed on NASDAQ. Adobe acquired it in 2016. Lopatecki stayed on at Adobe as Senior Director of Innovation before deciding it was time to build again.

The Arize Thesis: Observability Is Not Optional

The parallel Lopatecki draws between programmatic advertising and AI evaluation is instructive. In both cases, the underlying systems make millions of decisions per second using complex models. In both cases, the question is not just whether the model runs - it's whether it's making good decisions, whether its behavior is drifting, and whether the signals indicating failure arrive before users experience them. At TubeMogul he built reporting infrastructure that answered these questions for advertising. At Arize, he's building it for AI.

Arize's platform monitors AI models and LLM applications in production: detecting hallucinations, measuring drift, tracking evaluation metrics, and surfacing anomalies before they become incidents. The open-source tool Phoenix - built on OpenTelemetry standards so AI traces work like the infrastructure logs engineering teams already know - now has over two million monthly downloads. That number is not an accident. Making the foundational tooling free creates the ecosystem; the enterprise platform creates the business.

The Council of Judges

One of Lopatecki's specific contributions to how the industry thinks about LLM evaluation is the "council of judges" framework - using multiple AI models as evaluators of other AI outputs rather than relying on any single judge. The intuition is similar to how human review panels work in high-stakes settings: no single opinion should have unchecked authority. For LLM outputs that are probabilistic and context-dependent, triangulating across multiple evaluation signals reduces the noise.

This kind of operational thinking - borrowed from how reliable systems actually work in production - characterizes how Arize approaches the problem. It's not academic. The company's customers include Uber, Klaviyo, and Tripadvisor: organizations for which AI reliability is not a research question but an uptime question.

The $70M Series C and What Comes Next

In February 2025, Arize AI raised a $70 million Series C led by Adams Street Partners - described at the time as the largest single investment in AI observability. The round included M12 (Microsoft's venture arm), Datadog, and PagerDuty: companies with a direct financial stake in the reliability of AI infrastructure. That's not a coincidence. When a monitoring company and an incident-response company both back your observability startup, they are making a bet about where the pain is going to be.

Total funding stands at $131 million. The team is around 120 people. Arize:Observe 2025 introduced Alyx (an AI-native agent for evaluation workflows) and ADB, Arize's proprietary OLAP database engineered specifically for AI trace analysis at scale. The company is building toward a future where AI observability is as standard as application performance monitoring - something no serious engineering organization runs without.

Tinkering as a Practice

Away from the company, Lopatecki builds personal machine learning projects with a focus on unsupervised learning and deep neural networks. For someone running a 120-person AI company, staying close to the code is less unusual than it sounds - it's the difference between knowing what practitioners actually need and guessing from a distance. His blog posts on Arize's site range from hierarchical memory management in agent harnesses to analysis of Apple's paper on LLM reasoning limitations.

That last topic - what AI systems actually can and cannot do - is the thing Lopatecki keeps returning to. Not the hype, but the mechanics. Not what the press release says, but what happens when the system meets a real user with an unexpected input at 3am. That gap is why Arize exists. And it's why the company keeps growing.