▶ Watch ↗18:56
Conference affiliation: Citadel Securities
João Wenzel builds AI products around information that is difficult to collect, interpret or recover. His work has ranged from obtaining speech data for language education to building searchable podcast transcripts. At Citadel Securities, he works on firm-wide agent deployment and institutional memory: systems that distinguish what teams report about their projects from what the underlying evidence suggests.
João’s 2013 master’s thesis at Oxford examined how collaboration relates to Wikipedia article quality. He combined computational methods with user reviews, studying editing activity, distinct contributors and collaboration patterns. His professional profile also records his work on YouTube strategy. Andrew Ward, who worked with him for about three years, described João’s development from quantitative analysis toward a broader role advising the business, including work on the data and forecasts behind music negotiations.
At Google, João helped establish a team exploring AI for language education. Speech recognition performed much worse for strongly accented English, and obtaining suitable training data became a substantial part of the work. The team needed particular accents, spontaneous and scripted speech, and reliable transcriptions. João spent months assembling those inputs. In his later essay on specialized data bounty hunters, he proposed providers who would coordinate collection, cleaning and delivery against a buyer’s specifications, sparing each model-building team from having to organize the entire operation itself.
At Netflix, João led work applying machine learning to matching actors’ lip movements with dubbed dialogue. He subsequently worked on decentralized-storage adoption at Protocol Labs. His role as product lead for client growth included tools and process improvements for bringing large datasets onto Filecoin. These roles addressed different obstacles to adoption: a distracting viewing experience, missing training material and the difficulty of moving data onto unfamiliar infrastructure.
João announced that he had joined Catena Labs as chief product officer. His interests there included combining AI’s ability to interpret information with cryptographic mechanisms for provenance and incentives, while keeping the product focused on recognizable user problems.
Two independent projects show how João develops an idea into working software and examines where it falls short:
At Citadel Securities, João’s team applies state estimation to organizational knowledge. In robotics, a state estimator revises its understanding of the world as observations arrive. For a project, the corresponding questions are how old an observation is, where it came from and whether other evidence agrees with it. Indexing documents alone does not resolve conflicts between an optimistic status deck and a more recent discussion of failing tests.
In one example, a slide marked a project green while a comment identified an unresolved dependency. The system preserved the reported green status and proposed a yellow assessment, recording the reason for the difference. Keeping both accounts lets a manager investigate the disagreement rather than receive another summary that repeats the slide.
The team also changed how it processed the material. Giving a model the entire collection lost useful detail; focused specialist workflows, with other agents reviewing their work, produced more useful updates. Missing observations required separate treatment. A channel going quiet could mean that an information source had stopped working, rather than that a project had become inactive.
Historical records remain intact, allowing the team to test a warning using only information available when it was made. João reported that green projects flagged by the system were twice as likely to turn yellow or red in later records. He treated that result as a way to direct attention, rather than a prediction of precisely what would go wrong.
Agents propose revisions; code and stakeholders determine what enters the record. João distinguishes feedback that fixes one fact from feedback that teaches a rule applicable to future cases. Collecting that input in the tools people already use makes correction part of their work. The resulting record supports leadership discussions that begin where reported status and evidence disagree, and gives other agents context they can reuse.
▶ Watch ↗18:56
Loading more talks and highlights…