About Koodos Labs
Koodos Labs is a New York–based consumer AI research and product company with roots at Harvard’s Berkman Klein Center.
At Koodos Labs, we’re building the richest personal context model a user controls. Shelf, our core product, is a platform that brings together activity from a user’s connected services (from Spotify & Netflix to Amazon & ChatGPT) to build an evolving understanding of one’s preferences, habits and interests, and is a gateway to personal apps built by our community. Over time, external services will write to and query a user’s Shelf whenever they want to update memory or understand you better, moving us toward an internet that knows you and does things for you, on your terms.
The Personal Context
Shelf holds the digital footprints of all services that a user has chosen to connect. As an example, here is one person’s Tuesday as Shelf sees it:
A playlist in the afternoon. An episode after dinner. A diary entry for a film they saw at the weekend. A book added to a to-read list at midnight. Each of those is a row, and every row points to the same pseudonymous person. That is the unit of analysis, and very few datasets have it. Most studies of media and people have one platform, or one category, or self-reported preferences instead of behavior.
The scale: Shelf has millions of users, and tens of billions of events across more than a dozen platforms that Shelf offers as connections.
The individuals behind the rows are international, and they chose to connect several services and keep them connected. So the footprints are long and dense, not a one-off survey panel. And many of them take part in in-app studies.
In addition to the longitudinal cross-category digital footprints, Shelf has an in-app ecosystem of experiences, powered by AI agents, that users regularly return to. This is our environment where not only can we engage with users directly, but we can also test whether we are delivering value to users. In addition to the longitudinal consumption histories, Shelf has an ecosystem of native apps and agents - each with their own functional utility - acting as ample RL Environments that allow Shelf to understand if it is delivering on the user’s preferences and supporting the user with whatever they’re looking to achieve on that app.
Research areas of interest
We’re interested in the following questions, and welcome ideas beyond them:
Modeling and compression: What is the optimal way to compact someone’s digital footprints into a reliable, valid and useful personal context model?
Inference: What can cross-platform digital footprints reveal about preferences, intentions and future behavior? Do findings generalize and compound across platforms and/or medium?
Simulation: Can a model capture an individual’s unique responses rather than reproduce population averages? How do we validate simulations against actual behavior?
Evaluation and learning: Can AI agents autonomously propose better experiences for users that users benefit from and enjoy?
How we support research
Depending on the project, support may include:
Study design: Working together to define questions, study groups, baselines and evaluation methods.
Research environment: Access to a rich, longitudinal dataset of consumption behavior from millions of people across consumer platforms, as well as rich traces from across an ecosystem of apps & agents. Our cleanup pipelines and time cutoffs support reliable, reproducible evaluation.
Tooling and expertise: Documentation, analytics infrastructure and technical support.
Network: Collaborators can engage with our network of experts and researchers from across leading institutions, including Mozilla, Harvard, DeepMind, Cambridge and others.
Publication and visibility: Support for open-access papers, presentations and open-source work, and sharing findings on our blog. We welcome results that challenge our assumptions.
Interested?
We welcome researchers in machine learning, HCI, behavioral science, computational psychology, quantitative marketing and similar fields.
Send a short note to research@koodos.com with your background & research proposal. Early-stage ideas are welcome, and we’re happy to help discuss proposals before you submit them.