I build AI systems that help journalists and regulators document the networks behind online deception campaigns.
Executive Director and co-founder of SimPPL, a US 501(c)(3) research lab rebuilding digital trust, and a former postdoctoral affiliate at Boston University and MIT.
Platforms add warning labels, hide posts and let people block accounts. They almost never publish evidence that any of it works. I measure what happens after the message moves to other platforms, which the company cannot see. And I design market rules where a seller who lies loses money.
Academic job market paper Market Design Interventions for Safer Agentic AI (ICIS 2025). There is also a first-author working paper on time-series causal inference for social networks.
Reputation systems break when reviews are cheap to fake, and content moderation arrives after the sale. With Marshall Van Alstyne, Nina Mazar and Aaron Nichols at Boston University, I worked on truth warrants: a seller stakes money on a specific claim, and a buyer who was misled can collect. In my view that is the only lever here that scales, because it changes the incentive instead of policing the text. I helped found BU's Platform Governance Lab, now 25+ members. A companion paper, Certifiably True, is under review at Nature Communications.
Average profit per agent per game in a simulated two-sided ads market. Green bars are profit earned on truthful claims, red on false ones, and each cluster is one seller strategy. The left panel is a conventional ratings-and-reputation market; the right panel requires the seller to stake a warrant.
TakeawayThe generative AI seller running the deceptive strategy earns 16.62 per game under ratings alone and 9.30 once a warrant is required, a fall of about 44%. The bot making truthful claims is untouched, moving from 17.83 to 18.07. The mechanism prices deception without taxing the sellers who tell the truth.
My dissertation at NYU's Center for Social Media and Politics measured the causal effect of Twitter's soft and hard interventions on Trump's misinformation tweets during the 2020 election, advised by Richard Bonneau and Jonathan Nagler. My students have since taken the method to newer platforms: on Bluesky, issuing a block increases how much a user posts afterwards, and across 2M+ Truth Social posts a poster's follower count predicts amplification better than the quality of the news source does. Those three papers are on the publications page.
Cumulative retweets over the 24 hours after publication, split by what Twitter did: applied a warning label (soft intervention), removed the tweet (hard intervention), or left it alone.
TakeawayLabeled tweets end the day at roughly 47,000 retweets against 36,000 for comparable untouched tweets, so on-platform the label reads as a backfire. That descriptive gap is what motivated the paper. Following the same messages onto Facebook, Instagram and Reddit shows the intervention did reduce spread off-platform, which is the half a platform cannot see from its own dashboard.
I build multiagent AI systems that do the parts of an investigation a person cannot do at scale. One is a retrieval agent for conflict-region audio: it answers the analyst's question, highlights the clip it came from, and cites the source. Another is an agent-based model of Reddit I built at Oxford with Philip Torr and Atilim Gunes Baydin, restricted to the actions real users can take, which measures how coordinated inauthentic accounts change what the recommender surfaces. The write-ups are on the publications page.
Left: 4,500 Telegram channels spreading pro-Russian narratives about Ukraine, colored by cluster. Right: the subnetwork of those channels that promoted a conspiracy theory about the Moscow blasts.
TakeawayMany of the same channels that pushed the Ukraine campaign later promoted a conspiracy theory about the Moscow blasts. Showing which channels did it gives a reporter something to work with, because nobody can file a story from the whole network at once. We ran this investigation with former intelligence agency personnel.
A reporter keeps using the tooling after I leave. I design and run the AI capacity building for SimPPL's newsroom partners: DW Akademie in Germany, the NEST Center for Journalism in Mongolia, Africa Uncensored and Odipodev in Nairobi, and the Google News Initiative's JournalismAI Skills Lab. Arbiter was a JournalismAI case study in 2025.
Meta shut CrowdTangle down in 2024, and Twitter's API went behind a paywall the year before. Between them they supplied most of the public platform data this work depends on. The newsrooms we partner with collect and analyze their own data now.
A public observatory for digital discourse across six platforms. A reporter asks which local accounts are pushing election-fraud claims and gets a cited dataset back, method included.
The research lab I co-founded with Dhara Mungra to rebuild digital trust. Partners include Spreeha in Bangladesh, Migrasia in Hong Kong, Jagran New Media in India, VTDigger and New York Public Radio.
People across the global majority started asking AI chatbots their health questions before anyone checked the answers were safe. I co-founded Sakhi and built the retrieval behind it.
I came up through a non-premier engineering institute in India and started designing programs there while I was still an undergraduate myself, so students at institutes like mine would not lose out for want of resources, training and guidance.
The programs
Since 2018
I founded DJ Unicode as an open-source programming bootcamp, then Unicode Research for collaborative research, and in 2020 ran UMLSC, a 13-week machine learning course whose cohort became SimPPL. Together these have trained 70 to 100 students a year since 2018.
They grew into the NYU AI School, funded by DeepMind and Genentech, and the Foundations of AI Product Development course at the University of Mannheim, which my co-founder leads. NextGenAI is the current capstone fellowship. Write to me if you want one of these for your students.
Where the mentees went
Outcomes
Four of my mentees are doing PhDs in computer science, several are in graduate programs, and others work at FAANG companies and startups. Between them they have published 10+ papers at ICWSM, AAAI, NeurIPS and ICML, usually as first authors. I sit on three Boards of Studies at D.J. Sanghvi College of Engineering, and I mentor through Lumiere Education, Make a Difference and the UAlberta Career Mentoring Program.
What the programs won
Grants
Two Google exploreCSR awards (USD 32,000 and 75,000) and the inaugural Mozilla Responsible Computing Challenge in India (USD 25,000). Student testimonials are here, and they are the teaching evaluation I would rather be judged on.
Before all of this
Industry and CERN
I worked at Twitter on civic integrity, where I improved the precision of a civic and health misinformation classifier by 20% and scaled it to 1M+ tweets a week. Before that: multimodal recommender systems at Adobe Research (one US patent), product data science at Slack, Meta-funded work at Oxford, and CERN, where I built graph neural networks for particle track reconstruction and deployed the DeepJet package into production across 42 sites.
I strongly believe transparency begets trust, and that mechanisms for digital transparency have to come before any rebuilding of digital trust. I also do not think technology alone will solve this, which is why I joined the Board of a policy think tank. I have worked at platforms, outside platforms, and as an academic studying them, and that combination is most useful in a room where someone is drafting a rule.
Integrity Institute
Think tank
I sit on the Board of the Integrity Institute, a think tank of platform integrity professionals. Read the 2023 annual impact report. Through it I have published technical guidance on elections integrity and platform transparency, and consulted with UK and European agencies writing digital safety regulation.
Rooms and stages
Speaking
I speak regularly at TrustCon and the Stanford Trust and Safety Research Conference, and I have presented to technology regulators at the World Economic Forum. At the Atlantic Dialogues I represented the youth voice to an audience of presidents, ministers and leaders from the UN, World Bank and IMF. The UNDP has facilitated our presentations to several governments and embassies. I was a Responsible Tech Affiliate at All Tech is Human.
Recognition
Recognition
Atlantic Dialogues Emerging Leader, Google Research Innovator, CTS Belfer Fellow, ITS Rio Fellow, ISPI Next Business Leader, and Future Today Institute's People to Watch in AI and Local News. Economic Times Campus Stars listed me among India's top 33 engineers in 2018.
Where the work has been covered
Press
I studied art for 14 years and don't talk about it much. Graphite and ink, and some Procreate.
Four of the twenty-five drawings on the gallery page. The full gallery is here.