Research

Thoughtful Things: Local Reasoning for Smarter Devices

Our work on thoughtful things reimagines smart devices as collaborators capable of reasoning about their own behavior. Traditional smart systems rely on rigid rules or cloud-based intelligence, leaving them brittle, opaque, and dependent on connectivity. We instead design devices that can think locally, using compact, fine-tuned language models trained to reason about their internal state and generate both actions (“make it tropical”) and explanations (“I dimmed the light because you asked for a warm glow”).

This idea began with Sasha, a smart home assistant that uses large language models to interpret under-specified, goal-oriented commands such as “help me sleep better.” Sasha demonstrated that language models can creatively devise multi-device action plans, but also revealed reliability and privacy limitations when reasoning occurs in the cloud.

Our next step, building Thoughtful Things, moves that reasoning onto the devices themselves. We developed a five-stage framework that bootstraps local models entirely from synthetic data derived from a device’s formal state description. This approach allows even resource-limited hardware like a Raspberry Pi to host a self-contained language model that learns how to both act and explain without ever sending data to the cloud.

By uniting formal modeling and generative reasoning, thoughtful things make everyday devices more adaptive, interpretable, and privacy-preserving. They hint at a future in which technology not only responds, but can articulate why, bridging human understanding and machine autonomy in the spaces where we live and work.

Opportunistic Learning in the Internet of Things

Our work in decentralized federated and opportunistic learning explores how connected devices like smartphones, wearables, sensors, and other edge systems can learn from one another without ever depending on a central server or cloud connection. Instead of sending data away for training, these devices learn collaboratively through brief, peer-to-peer encounters, sharing models and insights directly.

This vision, which we call opportunistic collaborative learning, turns mobility and connectivity limits into assets: devices train locally, then exchange what they’ve learned when their paths cross. Over time, a community of devices collectively improves its models, even in settings where the network is unreliable or the data too sensitive to leave the device.

Our research spans the full ecosystem needed to make this approach viable, from algorithms that adapt to short, unpredictable encounters and heterogeneous hardware, to systems that simulate large-scale decentralized learning environments, and incentive mechanisms that ensure fair and trustworthy participation. The result is a new paradigm for privacy-preserving, resilient, and context-aware learning that can thrive entirely at the network’s edge.

Ultimately, this work imagines a future where learning is ambient and autonomous – where intelligence emerges not from a single powerful model in the cloud, but from the spontaneous, decentralized cooperation of everyday devices.

Education, Identity, and Belonging in Computer Science

Our research on computing identity and belonging explores how learners come to see themselves as computer scientists and engineers, and how environments, experiences, and communities help them get there. We study identity as something that can be designed for: through reflection, representation, and authentic participation, rather than as a byproduct of success.

Across K–16 settings, our work examines how pedagogical design and institutional structures influence persistence and engagement. In Project moveSMART, we integrate computational thinking into elementary physical education classes, connecting coding and data science with movement to broaden participation early and equitably. Our engineering outreach studies follow these pathways into adolescence, showing that both middle school campers and undergraduate camp counselors develop engineering identities and community cultural wealth through authentic, community-based learning experiences. At the collegiate and organizational level, our seed grant research examines how inclusive institutional practices – like funding students and staff as change agents – build community and a sense of belonging across the academic hierarchy.

Together, these studies reveal that identity and belonging are not peripheral to computing. They are the infrastructure of persistence. By designing learning environments that honor students’ assets, broaden access to computing concepts, and empower diverse voices to lead change, we aim to make computer science not just something students learn, but something they own.