Henry
Ehrenberg

Around 2016, most of the AI world was racing to build better models. Bigger architectures, more compute, flashier benchmarks. Henry Ehrenberg and a group of Stanford researchers asked a different question: what if the training data was the actual problem?

It wasn't a popular position. It wasn't even an obvious one. But inside Stanford's Hazy Research lab, under Professor Chris Re, Ehrenberg and his collaborators were building something they called Snorkel - a system that let users train state-of-the-art machine learning models without labeling a single training example by hand. Instead, users wrote small functions expressing domain knowledge. The system learned to combine them into probabilistic labels at scale.

The paper that emerged from this work - "Snorkel: Rapid Training Data Creation with Weak Supervision" - became one of the most cited works in data-centric AI. It didn't just describe a technique. It identified a gap that the entire industry had been ignoring while staring at leaderboards.

Before Snorkel AI had a name, it was a research problem. Ehrenberg was part of the original group at Stanford that developed the weak supervision framework, contributing to papers that would shape how the field thinks about data. His technical background - formal mathematics at Yale, computational engineering at Stanford, plus a stint doing representation learning at Facebook Applied AI - gave him range that pure ML researchers often lack.

At Facebook, Ehrenberg worked on the applied AI team's representation learning group, building systems that operated at the scale of one of the largest social platforms in the world. It was the kind of hands-on engineering exposure that most academic AI researchers never get. He brought that practical sensibility back to the startup.

When Snorkel AI incorporated in 2019, Ehrenberg took the Head of Engineering role alongside co-founders Alexander Ratner (CEO), Chris Re, Paroma Varma, and Braden Hancock. The founding team was unusual: more academic firepower than most enterprise software companies see in a decade, wrapped in a company that had to ship products on real deadlines for Fortune 500 clients.

Snorkel AI has three types of co-founders: the researchers who conceived the academic work, the operators who scaled the business, and the engineers who built what customers actually use. Ehrenberg sits squarely in that third category, serving as Head of Engineering and responsible for the technical strategy that keeps Snorkel's platform competitive in one of the fastest-moving markets in technology.

The engineering challenge at Snorkel isn't trivial. The company's platform needs to work for a compliance team at a bank, a clinical NLP team at a hospital system, a document intelligence team at an insurer, and a research lab training foundation models - each with different data schemas, security requirements, deployment constraints, and evaluation criteria. Building software that works across that range without becoming a mess of enterprise configuration switches is genuinely hard.

Ehrenberg's approach has been to stay close to the research side of the company while managing a production engineering organization. He has co-authored blog posts on enterprise GenAI evaluation, contributed to thinking about specialized evaluators as scalable proxies for subject matter experts, and maintained the academic rigor that set Snorkel apart from less principled competitors.