Stefania Druga

Stefania Druga is an AI researcher who created Cognimates, an open-source platform through which children program and train AI, and coauthored research on AI-supported creative programming with families. Her portfolio describes her as a Staff Research Scientist at Sakana.ai's RSI Lab working on self-improving research agents; that role statement is undated.[3][4][2]

In the June 2023 Latent Space discussion of OpenAI's Functions API, Druga described herself as a Microsoft Research intern also working with Fixie. Drawing on a Fixie hackathon she had run, she contrasted newcomers approaching agents through natural-language prompts with developers accustomed to Python and APIs. She asked how function calling would change onboarding and the balance between English and code, especially when users struggled to constrain or verify model outputs.[1]

Druga questioned how a model's choice of functions could be constrained through schemas describing their inputs and outputs, or through training that incorporated retrieval. Sharing and reusing functions, she pointed out, would create a larger search space. She expressed interest in knowledge graphs and retrieval-aware training that would offer alternatives to relying on vector databases. These were exploratory proposals, rather than descriptions of a system she had implemented.[1]

She argued for smaller, more efficient models, imagining on-device systems with greater privacy, security, and affordability, and less concentrated control over data and training. She also cited compute constraints and the need to reduce latency for interactive applications. Her interface critique was direct: “I don't wanna chat with everything.” Chatting separately with calendars, email, and other applications, she argued, would require a different approach to UI and UX.[1]

In the same discussion, Druga credited Leon and his team for the LLM security-probing tool Garak and raised a disclosure dilemma: publishing vulnerabilities and red-teaming examples could help people understand risks while also assisting people seeking to abuse models and their functions.[1]

Before these agent discussions, Druga had developed tools for children to learn how AI works by programming and teaching it. Her earlier work provides concrete examples of how she designed interactions around users with limited technical experience.[3][1]

In her October 2018 account, Druga recalled working for Google's Search Quality Team in Dublin for a year before leaving to pursue education more directly. She spent four months volunteering at an orphanage near Phnom-Penh, where older children passed what they learned on to younger children. She founded the STEM education organization Hackidemia in 2012. Later, believing its grassroots work was reaching its limits, she sought new skills and guidance through the MIT Media Lab's master's program; she entered without a specific project in mind.[3]

In early 2017, Druga and undergraduate intern Eesh Likhith built a Scratch extension for the Poppy Ergo Jr. robot, whose actuators could record and replay movements. They then prototyped a computer-vision extension using Clarifai's public image-recognition API. Their intended interaction was for a child to show an object to the robot's webcam and have it draw from the objects it already knew how to draw. They documented and published both extensions on ScratchX.[3]

Druga created Cognimates for children aged 7–14, building on Scratch, developed by the MIT Media Lab's Lifelong Kindergarten group, and the Personal Robots group's prior work on “programming as teaching” under Cynthia Breazeal. The platform combined visual programming, modular cognitive-service plugins, and tools for training AI models with images and text. Children could customize devices such as Alexa and Cozmo or build projects such as a Rock Paper Scissors game. Druga collaborated with Randi Williams on early child-and-parent AI research and credited Sarah T. Vu, Tammy Qiu, Clemente Ocejo, Eesh, and Lauren Oh with technical and field-research contributions to Cognimates.[3]

Children participated as design partners throughout development. Druga recalled that they called a sentiment-analysis activity the “make Nary happy” project, encouraging the team to use characters and stories to make abstract AI services visible. Nary expressed emotions when the computer detected happy or sad messages; a giant eye showed what the vision extension recognized or where it became confused. The account credits Mircea Dragoi of Lateral with Nary's design and character iterations. Druga reported that her thesis research with 107 children in four countries found improved understanding of AI concepts and changed perceptions of smart agents after programming and teaching Cognimates.[3]

Druga and Amy J. Ko coauthored AI Friends: A Design Framework for AI-Powered Creative Programming for Youth, submitted to arXiv on May 17, 2023. They built a Wizard of Oz platform: a researcher operated the AI Friend rather than leaving it to act autonomously. Across three weeks of programming activities with ten children aged 7–12 and nine parents, the authors reported that the Friend's questions made generating game ideas easier, while parents helped guide more complex programming tasks when the Friend failed to help. They argued that AI-supported platforms should emphasize children's agency and creative self-efficacy within family interactions.[4]