Amrit Puhan

I am a Software Development Engineer II in the Applied AI Solutions org at Amazon Web Services (AWS) in Seattle, building video search over petabyte-scale autonomous driving data and time series anomaly detection on industrial data.

Alongside my work at AWS, I do research in computational social choice, information retrieval and machine learning, driven by one question: how do you find the right answer when most people get it wrong? I study voting methods that recover the true ranking from noisy, disagreeing crowdsourced judgments, even when experts are in minority. This work began with my MS in Informatics at Penn State in the FAIR Lab, where I was advised by Dr. Hadi Hosseini and mentored by Dr. Debmalya Mandal. My thesis, Recovering Ground Truth Rankings When the Majority Is Misinformed, won the Best MS Thesis Award in Artificial Intelligence, and the work was published at NeurIPS, WWW and KDD. Before that, I earned my B.Tech in Computer Science and Engineering from NIT Rourkela, doing computer vision research with Dr. Anup Nandy.

my career at a glance

academia

Standard voting rules work when most people are more likely right than wrong. When experts are in minority, simply aggregating opinions, by any rule, cannot recover the truth. My research builds voting methods that still recover the true ranking in this setting. As a graduate researcher in the FAIR Lab at Penn State, I worked on three questions.

• Can the true ranking over many options be recovered when each person judges only a few? We designed methods that do, and proved that the number of votes needed depends on how many options each person sees, not on the total (NeurIPS 2024).
• When is recovery possible at all, and from how many votes? We modeled human voters as groups of experts and non-experts, and derived the conditions on each group’s size and expertise under which the true ranking is provably recovered, and how many votes it takes (WWW 2025).
• How should methods for this problem be benchmarked? We released our crowdsourced study as a public dataset with benchmarks (KDD 2026).

My work is equally theoretical and empirical, and I am now bringing it to multi-agent LLM systems around two questions.

• How should LLM agents reach a consensus answer when most of them are wrong? I am studying how to aggregate the answers of multiple agents at inference time.
• What labels should reward models be trained on when the labelers disagree? I am studying how to aggregate preference labels for reward models used in reinforcement learning based fine-tuning.

As an undergraduate researcher in the Intelligent Computing and Computer Vision group at NIT Rourkela, I worked on two questions.

• Can cognitive state be predicted from how people walk? We calibrated eight Kinect sensors to capture gait (CVIP 2021), reconstructed the scene in 3D, and used deep learning models to predict cognitive state from gait.
• Can webcam hand gestures work as passwords? We built a hand-gesture based password recognition system (ICCIS 2019).

selected publications

^* Lead author; authors listed alphabetically.

academic service

• Reviewer, internal Amazon science conference
• Reviewer, KDD 2026, Datasets and Benchmarks Track
• Reviewer, The ACM Web Conference (WWW) 2025, Main Track

industry

At AWS, I build AI over petabyte-scale automotive and industrial data: multi-modal video search that helps ADAS/AV developers find rare edge cases with natural language queries, and time series anomaly detection for predictive maintenance. Before AWS, I built front-end applications for SAP SuccessFactors at SAP Labs India, and a door-entry system at Schneider Electric that could be unlocked from phones, laptops and other devices on any operating system.