Bryan Bischof is Head of AI at Theory Ventures, where he investigates whether AI agents can perform demanding analytical work reliably. A mathematician and longtime machine-learning practitioner, he previously led development of Hex Magic, embedding AI assistance throughout the data-analysis platform.
Bischof earned a PhD in pure mathematics from Kansas State University and worked across IBM, Blue Bottle Coffee, Stitch Fix, Weights & Biases, and Hex. As Blue Bottle’s first data hire, he built a team addressing data warehousing, demand forecasting, analytics, and recommendations. He later worked on data science at Stitch Fix and established data science, machine-learning, and data-engineering teams at Weights & Biases before leading AI at Hex.
At Hex, Bischof built tools that generate and explain SQL and Python, create visualizations, and break unwieldy queries into interconnected notebook cells. He favored notebooks as an interface for AI-assisted analysis because editable cells preserve analytical structure and keep assistance within existing workflows. His team abandoned a promising standalone experience when it threatened to pull users away from Hex’s core product.
Bischof and Hex engineer Grant Marvin reduced AI-generated chart latency from 10.3 seconds to 1.3 seconds by simplifying chart specifications, limiting generated tokens, constraining outputs, and testing faster models against explicit quality thresholds. With Hector Yee, he coauthored Building Recommendation Systems in Python and JAX; he also contributed to practical production guidance for language-model applications alongside Eugene Yan, Charles Frye, Hamel Husain, Jason Liu, and Shreya Shankar.
- Production-grounded AI evaluation: Bischof builds evaluations close to real application behavior, converts complex tasks into concrete checks, and monitors regressions when model providers change underlying systems. His checkpointed agent evaluations divide longer workflows into sequential tests that expose where analytical agents fail.
- Retrieval as a recommendation problem: Effective retrieval requires selecting and ranking context for a specific task. At Hex, Bischof paired LanceDB with customized retrieval to reduce hallucinations in analytical workflows.
- AI inside existing workflows: Bischof treats AI as infrastructure for improving familiar tools. His writing on agent skills as shared operational knowledge extends that principle to reusable expertise encoded by domain specialists.
- Realistic benchmarks for data agents: At Theory Ventures, Bischof helped organize America’s Next Top Modeler, evaluating agents against structured tables, logs, and documents. His analysis of trustworthy analytical agents warns that optimizing visible benchmarks can reward evaluation gaming over dependable performance.
Bischof also teaches in Rutgers University’s Master of Business and Science program, where his AI engineering course, first offered in spring 2026, addresses language-model applications and production constraints.
