Bryan Bischof

Bryan Bischof is Head of AI at Theory Ventures. Previously, he led Hex’s AI team building Magic, a suite of AI capabilities integrated into its data-analysis platform. His decisions there included rejecting fine-tuning specifically to teach models Hex’s notebook format, developing evaluations from real requests, and keeping generated work editable within the user’s workflow. He also coauthored Building Recommendation Systems in Python and JAX with Hector Yee, published by O’Reilly in December 2023.[3][1][5]

In his November 2023 Latent Space interview, Bischof described Magic as enhancements to an existing platform. Hex already had AI-assisted SQL and Python generation and code explanation when he joined. Its February 2, 2023 changelog listed five tools embedded in SQL and Python cells: Generate, Edit, Fix, Explain and Complete. Bischof authored the May 4 public-beta announcement; by October 5, Hex’s release introduced Magic Analysis to kickstart an analysis with AI. He also used Magic for his own analytics while making product decisions.[1][8][9][10]

Bischof rejected fine-tuning a model simply to understand Hex cells. Instead, the team abstracted notebook wrappers and chart representations away from the model, letting it generate familiar languages such as SQL and converting the output into Hex’s format. He connected this decision to earlier experience with GPT-2, GPT-3 and fine-tuned models, where preprocessing and postprocessing could avoid more extensive model work.[1]

His preference for general models was tied to Hex’s varied customers. A model trained on notebooks still needed to understand the business being analyzed: coffee, pastries and demand forecasting at Blue Bottle, for example. In November 2023, he reported using GPT-3.5 Turbo and GPT-4 in production, with GPT-4 required for many applications. His argument concerned the broad domain knowledge needed by an analytics platform.[1]

Bischof treated notebook cells as atomic ideas that an AI assistant should help users manipulate. He described the Hex team’s June and July work on an internally named “airlock”: a sequence of related generated cells that users could preview, then accept or reject together. He called notebooks “chat plus plus” and preferred “AI in the loop” as a design principle: the user stays at the center, reaches for AI as a tool, and remains in charge of the workflow.[1]

Bischof’s formulation was that “ultimately RAG is just RecSys for LLM.” For Magic, retrieval had to supply resources that a general model could not be expected to know: the customer’s warehouse structure, relevant tables, joins and data caveats. A request about customer trends might otherwise produce plausible SQL against invented tables while missing necessary joins or filters. He described improving Magic’s retrieval as building a recommendation system for the resources an LLM needs to execute a task.[1]

He drew on Eugene Yan’s writing about multistage recommenders and Even Oldridge’s related work on Merlin to explain the available engineering levers. Retrieval, ranking and serving each offer ways to improve a system. In clothing recommendations, serving includes enforcing diversity, relevance and stock constraints; analogous downstream checks can improve LLM outputs even when engineers cannot retrain the underlying model.[1]

Bischof recalled starting Hex evaluation work in February 2023 and manually reviewing internal Magic requests to turn them into tests. He argued for separate checks alongside overall task success: whether SQL references a real table, whether a generated variable collides with one already in scope, or whether a completion leaves a brace unclosed. These checks could expose failures that engineering could detect and repair. His team also included an AI engineer focused on experiments—testing hypotheses about model behavior and extending evaluations.[1]

In “Making AI charts go brrrr”, published June 20, 2024, Bischof and Grant Marvin explained the Hex team’s insistence that AI generate native, interactive chart cells. Users needed to evaluate and edit generated work with the same tools they used manually. The constraint complicated implementation: GPT-4 could generate Vega-Lite, but that representation was verbose, while Hex’s smaller internal chart specification was unfamiliar to the model.[4]

The team reduced the output specification to a “chart stub” and handwrote 100 evaluations using Spellgrounds, Hex’s evaluation framework designed to run scenarios close to production. Bischof and Marvin reported that a 35-feature stub with GPT-4 Turbo produced correct responses 90% of the time in under five seconds. Switching to GPT-3.5 made generation faster but reduced accuracy to barely better than a coin flip. Production usage then showed that roughly 90% of generated charts needed only 10 of the 35 features. A smaller stub, with enumerated choices for three features, brought GPT-3.5 to 80% accuracy at around one second in their evaluations.[4]

The authors reported reducing average chart-generation time from 10.3 to 1.3 seconds. They also described a speed improvement they declined: removing chart and axis titles reduced streaming time to 0.7 seconds but lowered accuracy by eight percentage points to 72%, below their quality bar. They kept the titles at a cost of approximately 0.5 seconds. Their explanation—that generating titles first might help guide the remaining specification—was a hypothesis.[4]

At Theory Ventures, Bischof’s work includes the management of reusable agent skills. In “Getting Over Your Skills Issues”, published in June 2026 and updated in September, he argued: “Skills are software whether you like it or not.” Instructions acquire ownership, dependency, versioning and governance problems as they become shared and connected to tools.[7]

Bischof organized skill complexity along two axes: capability, or what a skill can access and do, and distribution, or how widely it travels and who depends on it. He recommended infrastructure proportional to those needs: local files for personal instructions, versioning and possibly review for evolving shared skills, and a system of record for shared skills that adapt automatically. He described Theory’s team administering some skills through Claude Enterprise, primarily packaging them with its internal MCP deployed through Prefect’s Horizon, and building middleware to record usage in MotherDuck.[7]

Before leading AI product work, Bischof worked in pure mathematics, retail data systems and recommendation systems. His earlier coffee and clothing work supplied concrete examples for his later explanations of domain knowledge, retrieval and production constraints.[6][1]

Bischof graduated from Westminster College in 2008 with a mathematics degree and entered Kansas State University’s mathematics PhD program. In his January 2024 alumni interview, he recalled six years working in areas related to algebra, geometry and mathematical physics. His advisor died during his fifth graduate year, extending his timeline by a year. He said financial constraints led him into industry despite his love of research, initially working on time-series processing at a recently IBM-acquired startup in California. He earned his PhD in pure mathematics.[6][3]

At Blue Bottle Coffee, Bischof was the first data hire and began by putting the data warehouse in order. He led and expanded the team, focusing principally on demand forecasting, alongside recommendations, website optimization and analytics. He described this as his first management role and recalled building a coffee recommender based on users’ favorite food flavors. He joined Stitch Fix after Blue Bottle’s Nestlé acquisition.[1][6][3]

At Stitch Fix, Bischof worked on machine learning, optimization, logistics and recommendation systems. He credited Reza with introducing him to GPT-2 and recalled considering clothing-description generation using existing clothing features; he also credited Stitch Fix’s NLP scientists working on applications of that kind. He subsequently built data teams at Weights & Biases. In his alumni account, he described leaving a later startup role to focus on AI research at an institute in Berkeley before joining Hex.[1][6][3]

Bischof teaches data science and analytics in Rutgers’ graduate school. In January 2024, he described himself as an adjunct professor and credited Dr. Christie Nelson with the connection that led to the opportunity. He emphasizes objective framing: students should be able to state an objective function and explain how it relates to the problem they are trying to solve.[3][6][1]

Bischof encourages broad experimentation with AI tools while teaching their limitations. In November 2023, he recalled opening the first class of his semester with GPT demonstrations showing both successful and unsuccessful data analysis. His January 2024 advice paired immediate experimentation with mathematical curiosity: people can develop AI applications without deep mathematics and statistics, but those foundations help them understand where applications may fail.[1][6]

His enthusiasm for applications came with skepticism about infrastructure hype. In 2023, he borrowed Adam Azzam’s “iron mine” analogy to stress the work required to turn AI capabilities into value. He was particularly excited by a personal “Memex”: a memory extender indexing books, podcasts and other information for natural-language retrieval. He also wanted more experiments connecting AI to physical hardware and daily life.[1]