A joint fellowship from Mozilla and koodos labs | Remote | Summer 2026 (10–12 weeks)
koodos labs is a consumer AI research and product company dedicated to ensuring that the internet knows you and can do things for you, on your terms. We are the company behind Shelf, the leading personal context platform used by millions of people to track and understand what they consume. Read about the co-founders and their story here.
Mozilla.ai is the Mozilla's new AI company building open source AI that works for people, not on them. Led by Prof John Dickerson.
AI models get evaluated on everything except the thing that matters most for truly personalized AI: do they actually understand you? There is no standard benchmark for how well a system can build a coherent model of a person from the full stack of user signals. Nobody has formally characterized that structure. So nobody can measure progress against it. And because the modeling happens inside closed platforms, the people being represented have no visibility into how they are represented.
No agreed taxonomy. No methodology for how signals compound into a holistic representation of you.
Your primary deliverable is a publishable benchmark: a rigorous, reproducible eval framework for user understanding across signal types.
Part 1: Formalize the structure of a user profile. Characterize the taxonomy of signals that make up a person's identity to an AI system. Build on our current taxonomy to test the structure's coverage and accuracy against real user data; Shelf has millions of users with aggregated cross-category data. Identify where it over-specifies, where it under-specifies, where inferences systematically fail. Ultimately, this’ll become the core representation of a user’s identity and taste to AI systems.
Part 2: Design and measure gap-aware collection. A structure that tells you which data would be most valuable to collect next drives a compounding loop. Identify gap, surface a prompt, collect the missing signal, update the profile, repeat.
Deliverables: a paper that can be built on for publication targeting RecSys, UMAP, NeurIPS, or adjacent; and an open-source contribution for a user understanding benchmark.