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 human-AI complementarity and oversight, epistemic integrity and fact-checking with LLMs, vulnerabilities and emergent behavior in multi-agent AI systems, interpretability, and AI for public health and medicine.
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.
What I'm working on and thinking about right now.
How do people respond to the confidence LLMs express? We find people recognize when a model's assertive language exceeds its actual confidence, yet still reward it, a tension we call confidence misalignment. I also contribute to Veracity, an open-source fact-checking assistant, and its browser extension, and study why fact-checking evaluation must separate ambiguous claims from genuine system failures.
A benchmark testing whether AI agents can identify vulnerability chains that span realistic enterprise systems (application code, secrets, infrastructure-as-code, cloud/IAM, network policy, and agentic API trust) rather than a single vulnerable program.
As AI systems increasingly act as groups of interacting agents, risky coordination can emerge that no single agent was designed for. CollectiveScope detects the onset of emergent collective behavior, identifies which agents drive it, and causally validates that attribution through targeted interventions.
Do independently trained ECG foundation models learn shared, reproducible representations of the same clinical concepts, or do apparent "clinical features" depend on architecture, seed, dataset, and method? Using sparse autoencoders and causal tests, I also ask whether a clinical concept is partly a demographic one.
A conversational partner that reads a proposed negotiation strategy and its supporting documents, surfaces loopholes, inconsistencies, and vulnerabilities, and suggests stronger alternatives, built to support human judgment rather than replace it.
A study of whether AI assistance helps CS course graders work more accurately and consistently, and how students perceive and respond to AI-generated feedback on their work.
A bilingual, two-center study evaluating local LLMs for extracting infectious disease syndromes from multilingual EHRs, plus an LLM tutor that generates cases and teaches through Socratic dialogue, piloting in Harvard's HST40 in Fall 2026.
A searchable, reaction-aware library and screening method for connecting terminal fragments into synthesizable molecules that bind multiple targets.
How did the rush of researchers from other fields reshape pandemic science? Across 450,000+ COVID-19 papers in OpenAlex, outbreak scientists never exceeded 12.5% of authors and 62.7% of papers had none on the team. Yet expert-led teams published in higher-impact venues and drew more citations, a pattern of expert dilution that fortified knowledge silos.
A quantitative framework for diversifiable risk in AI infrastructure investments, scoring site-level financial opacity, geographic vulnerability, resource stress, and community impact and aggregating them 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.