Steering Long-Horizon Agents
How can we build user models that better align long-horizon research agents with the people directing them? Ongoing at AI2.
PhD’ing at MIT CSAIL, graduating in 2027. I build methods that help data agents elicit, refine, and act on human preferences. During my PhD, I’ve been fortunate to work with Arvind Satyanarayan at MIT and collaborators at Ai2 and Salesforce Research. Previously, I built large-scale data systems at Microsoft.
I have given talks at PositConf, SciPy, and a variety of industry research groups. As a (literal) side quest during my PhD, I conduct polar research in Antarctica and the Arctic as a National Geographic Explorer.
Some of the recent things on my mind:
Embedding New Interactions in the Tools People Already Use
Help usually lives somewhere else: a separate app, a chat window, a course. I build interaction techniques that sit inside the interfaces people already use, so support arrives in the context of the work. A semantic reader lets screen-reader users navigate a page through a generated outline of its meaningful sections, with pointers back to the underlying content. We are post-training a local model to generate these outlines, so page content never has to leave the device, in co-design with Blind and Low Vision collaborators. Embedded learning systems bring deliberate practice into everyday work, attaching coaching and feedback to the doc, notebook, or spreadsheet someone is already working in. semantic reader blog post, embedded learning essay
Mapping an Evolving Analysis
Our project Ptolemy provides an interface for navigating and steering long-running analytical investigations. Ptolemy turns individual analytical moves into persistent, executable objects called analets, then uses CellQL and semantic embeddings to organize them into a navigable map. The map helps analysts understand where an inquiry has been, how different explorations relate, and where new analyses fall within the larger investigation. Work done at Salesforce Research with Vidya Setlur and Denny Bromley. paper. Patent Under Submission
Composing Interactive Instruments to Express Analytical Intent
Analytical intent is often grounded in specific data points, outliers, clusters, and relationships that are difficult to express in words alone. Meros is a declarative language for generating bespoke interactive visualizations through which analysts can express those intentions directly. Meros offers a concise set of components and a binding algebra that defines how they compose, letting authors build interactions that group, rank, and relate elements—restructuring an analysis rather than merely adjusting its parameters. Work done with Arvind Satyanarayan. docs video