PhD Student in Computer Science
Secondary Field in Mind, Brain, and Behavior
Harvard University
I'm a PhD student at Harvard, co-advised by Finale Doshi-Velez and Maimuna Majumder. I also work with Reihaneh Rabbany and Jean-François Godbout in the Complex Data Lab at Mila.
I study how AI systems can support human judgment and decision-making in high-stakes settings, and what happens when they fail: when models hallucinate, are manipulated, or communicate confidence they shouldn't have, and how people respond. My work spans multi-agent AI safety (vulnerabilities and emergent behavior), interpretability, human-AI complementarity and oversight, epistemic integrity and fact-checking with LLMs, and AI for health.
Before Harvard, I earned a B.S. in Bioengineering from UC San Diego (valedictorian, summa cum laude) with minors in Data Science and Cognitive Science, where I did research in soft robotics and in machine learning for particle physics.
Outside the lab, I love trying new things, from figure skating and climbing to crochet, baking, and bird watching on the trail.
More about meWhat I'm working on and thinking about right now.
Multi-agent AI systems can develop coordinated behavior that no individual agent was designed to produce. CollectiveScope aims to detect when such collective behavior emerges, identify which agents drive it, and validate these attributions through targeted interventions.
A benchmark that evaluates whether AI agents can trace vulnerability chains spanning multiple components of an enterprise system (application code, secrets, infrastructure-as-code, cloud/IAM, network policy, and agentic API trust), rather than spanning a single vulnerability or single part of an application.
Do independently trained ECG foundation models learn the same clinical concepts? Using sparse autoencoders, I extract the features these models learn and test whether they are consistent across architectures, random seeds, and training datasets.
How do people respond to the confidence LLMs express? We find that people can tell when a model's assertive language exceeds its actual confidence, yet still favor those responses, a gap we call confidence misalignment. I also contribute to Veracity, an open-source fact-checking system, and its browser extension, and develop evaluation methods that account for ambiguity in the claims/evidence being checked.
An AI sparring partner that takes in a proposed strategy for a case, and interacts with the user (negotiator) to examine supporting documents, identify assumptions, loopholes, or inconsistencies that be surfaced and suggest stronger alternatives. The aim is to prepare negotiators with the skills needed for effective negotiation.
A study of whether AI assistance improves the accuracy and consistency of grading in introductory CS courses (focusing on Harvard's CS20 - Discrete Mathematics for Computer Science course), and how students perceive and respond to AI-generated feedback.
Do clinical LLMs perform as reliably at one hospital as at another? In a two-center study at Brigham and Women's Hospital and Amsterdam UMC, we run the same open-weight models on-site on each hospital's patient notes to identify lower respiratory tract infections (pneumonia, influenza, COVID-19, and RSV), using patient microbiology results as ground truth. The study measures where these models generalize across health systems and where further validation is needed before clinical use.
An LLM tutor that generates infectious disease cases and teaches through Socratic questioning rather than direct answers. Students practice interviewing a simulated patient, work toward a diagnosis, and review the relevant microbiology. The tutor will be piloted in Harvard's HST40 (Mechanisms of Microbial Pathogenesis) course
A reaction-aware fragment library and screening method for designing synthesizable molecules that bind multiple targets by linking terminal fragments.
How did the influx of researchers from other fields shape COVID-19 science? Across 450,000+ papers in OpenAlex, outbreak scientists never made up more than 12.5% of authors, and 62.7% of papers included none. Papers with outbreak experts appeared in higher-impact venues and received more citations, suggesting that expertise was diluted across the field while expert-led work remained siloed.
A quantitative framework for assessing diversifiable risk in AI infrastructure investments. It scores sites on financial opacity, geographic vulnerability, resource stress, and community impact, and aggregates these scores into portfolio-level exposure.
Interested in collaborating? I'm always happy to talk about research ideas, especially around human-AI decision-making, AI safety, and AI for health. Reach out at sukanyakrishna@g.harvard.edu.
* denotes equal contribution. Full list on Google Scholar.