Senior Research Scientist @ Salesforce AI Research

Pranav N. Venkit

Building AI agents that are trustworthy by design and aligned with the people they serve.

Hi, I'm Pranav (PRAH-nuv). I work on trustworthy agentic AI: safety and privacy evaluation, post-training, persona-based decision making, and agent-to-agent communication, so AI systems earn human trust before they reach real people.

I
Portrait of Pranav N. Venkit
Trust & Safety Deep Research Agent Pipelines Agent-to-Agent AI Simulation Human-Centered AI
Scroll

Six domains, one constellation

My research space.

I've always loved space: mostly unexplored, with a few bright points you learn to connect. So I drew my research the same way. Every star is a paper, placed between the domains it bridges, with Trust & Safety at the core because it runs through all of them. Press play to watch five years take shape.

Highlights2026
01 Invited lectureMLSS 2026Columbia University, NYC

Evaluating AI Agents in the Era of Benchmarks

Taught ~200 PhD students at the Machine Learning Summer School how to rigorously measure, benchmark, and stress-test agentic systems as they take on complex, real-world tasks.

02 TutorialKDD 2026Jeju, Korea

Agents in the Wild: Where Research Meets Deployment

A tutorial on what changes when agents leave the lab, co-presented with G.H. Yang, H. Sedghamiz, E. Santus, V. Dibia, and I. Baldini.

About

I study what AI agents do once they leave the lab (what they cite, what they leak, how they persuade each other, and who they claim to speak for) and build the evaluations and training that make them worth trusting.

I'm a Senior Research Scientist at Salesforce AI Research on the Interactive AI team, where my work sits at the intersection of agentic systems and trust and safety. I build safety and privacy evaluations for agents that act on real data: when they overstep, leak sensitive information, or break policy, including in the messages agents send each other. I then close the loop with post-training, turning those failures into data that teaches models to behave better, and training reasoning models to propose diverse, grounded scientific ideas (Diverse Scientific Ideas, 2026).

A second thread is persona-based decision making. LLM personas increasingly stand in for real users and make choices on their behalf, so I ask who they actually represent (A Tale of Two Identities, AAAI 2026), how to ground them in social science (InterviewSim, COLM 2026), and how they drift out of character over long conversations. I also study agent-to-agent communication: how persuasion and beliefs cascade through agent networks (Belief Cascades) and how images sway vision-language agents (MMPersuade and Seeing is Believing, EMNLP 2026). My deep research work audits answer engines for unsupported claims (DeepTRACE, ICLR 2026; FAccT 2025) and builds research agents that know when to stop (ACL 2026).

All of it is human-centered AI at heart, and it grows out of my Ph.D. in Informatics at Penn State with Dr. Shomir Wilson, where I studied sociodemographic bias in NLP across disability, nationality, race, and culture. That work earned an Outstanding Paper Award at EMNLP 2023, a Best Short Paper Award at TrustNLP, a DEI Recognition Award at CSCW 2024, and an Honorable Mention for Penn State's Best AI Dissertation. I serve as an Area Chair for ACL and EMNLP and on the program committees of FAccT and AIES.

Full background Google Scholar ↗
As seen in
Reach out

Reasons to say hello.

Thanks for visiting! Here are a few good reasons to get in touch. The list is probably incomplete.

You're a researcher

Questions about a paper, an idea you'd like to brainstorm or debug, or a possible collaboration on agent evaluation, deep research, personas, or responsible NLP.

You're a student

Thinking about a Ph.D. in NLP or HCI, or about doing research in industry? I'm glad to share what I've learned, especially if you're making the jump from engineering like I did.

You're a journalist

Looking for a technical perspective on AI search and deep research, LLM personas, or bias in language technology. My work has appeared in New Scientist, Fast Company, and The Hill.

You're building agents

Shipping an agent or LLM feature and want to think through evaluation, safety, or user simulation before it meets real users.

You're organizing something

A talk, panel, tutorial, or guest lecture on trustworthy agents or sociotechnical NLP.

You're in the Bay Area

Up for coffee and a conversation about NLP, HCI, or responsible AI with like-minded folks? Let's meet.

When in doubt, reach out!