CV
Contact Information
| Name | Amrit Puhan |
| Professional Title | Software Development Engineer II, AWS |
| amrit.puhan@outlook.com |
Professional Summary
Interested in how reliable collective judgments can be drawn from noisy, disagreeing people, and how that question carries over to evaluating and aligning AI systems with human preferences.
Experience
-
2026 - present Seattle, WA
Software Development Engineer II
Amazon Web Services, Applied AI Solutions
- Build AI systems over petabyte-scale automotive and industrial data at sub-second latency.
- Develop a multi-modal video search system for autonomous-driving (ADAS/AV) teams, helping developers surface rare edge cases across video, sensor and annotation data with plain-English queries.
- Develop time series anomaly detection for predictive maintenance of industrial equipment, evaluating supervised and unsupervised statistical and deep learning models.
-
2024 - 2026 Seattle, WA
Software Development Engineer
Amazon Web Services, Applied AI Solutions
- Contributed to an IoT service connecting vehicle fleets to the cloud for over-the-air software updates and data transfer.
-
2023 - 2024 State College, PA
Graduate Researcher
FAIR Lab, Penn State
Advised by Dr. Hadi Hosseini; co-advised by Dr. Debmalya Mandal (University of Warwick).
- Designed scalable Surprisingly Popular (SP) voting algorithms that recover ground-truth rankings from partial preferences when experts are a minority. The best variant (Partial-SP) recovers the true ranking about 50% more accurately than Copeland on votes alone (about 90% vs 60% of pairs ordered correctly). (NeurIPS 2024)
- Ran a crowdsourcing study with 432 participants on Amazon MTurk across three domains (geography, movies, paintings) of 36 alternatives each.
- Modeled voter behavior as concentric mixtures of Mallows and Plackett-Luce models, inferred parameters with Bayesian MCMC (No-U-Turn Sampler), and derived when SP voting recovers the ground truth with high probability. On real data, SP voting made about 83% fewer ranking errors than Copeland (Kendall-Tau distance about 0.8 vs 4.8). (NeurIPS 2024, WWW 2025)
- Released SP-Rank, the first large-scale public dataset pairing personal votes with predictions of others’ votes. It holds 12,384 datapoints from 1,152 participants across 9 elicitation formats, open-sourced with benchmarks showing SP voting recovers rankings with a Kendall-Tau correlation about 400% higher than Copeland on votes alone (about 0.49 vs 0.10, averaged across formats). (KDD 2026)
- Used SP voting to generate preference labels for RLHF reward-model training, with weak and strong language models as voters, improving label accuracy by 9% on average over standard voting rules in preliminary experiments. (MS thesis)
- Fine-tuned LLMs for multiple-choice QA with LoRA, quantization and PEFT, ranking 6th on the AI2 SWAG leaderboard with 90.98% accuracy (2024).
-
2023 - 2024 State College, PA
Graduate Assistant (part-time)
Penn State Organizational Effectiveness
- Built Power BI enrollment dashboards for Penn State World Campus and a reusable template that cut development time by 60% and improved tracked enrollment KPIs by 25%.
-
2023 State College, PA
Instructional Assistant (Spring 2023)
Penn State College of IST
IST 261: Application Development (Java, MVC)
- Coached a 20+ student undergraduate class in object-oriented design, MVC architecture and debugging through office hours and one-on-one assignment help.
- Graded assignments with constructive feedback and served as a liaison between students and faculty.
-
2020 - 2022 Bengaluru, India
Associate Software Developer
SAP Labs India
- Developed five front-end applications for SAP SuccessFactors, an enterprise B2B cloud application.
- Quality Analyst for six projects; coordinated the product-wide quarterly testing event twice.
- CI/CD contact for ten teams; automated Jenkins deployments of 40+ applications across five countries.
-
2019 - 2020 Rourkela, India
Undergraduate Researcher
Intelligent Computing and Computer Vision Group, NIT Rourkela
Advised by Dr. Anup Nandy.
- Proposed two calibration methods for an 8-sensor Microsoft Kinect v2 gait-capture setup, reaching a mean reprojection error of 0.42 pixels, a 76.27% reduction over existing methods. (CVIP 2021)
- Predicted cognitive state from calibrated gait features with ML models (R² = 0.9556), suggesting gait is a strong indicator of cognitive state.
- Co-developed a hand-gesture password recognition system that runs on an ordinary laptop webcam, reaching 96.16% test accuracy and outperforming prior webcam (93.75%) and Kinect-based (86.5%) systems. (ICCIS 2019)
-
2019 Bengaluru, India
Summer Research Intern (May to Jul 2019)
Schneider Electric India
- Built a cross-platform IoT door-entry system (web, Android, iOS) with real-time text, audio and video, cutting latency by 83.3%.
Education
-
2022 - 2024 State College, PA
The Pennsylvania State University
MS, Informatics (Data Science concentration), GPA 3.88/4.0
- Advised by Dr. Hadi Hosseini.
- Thesis: Recovering Ground Truth Rankings When the Majority Is Misinformed. Best MS Thesis Award in Artificial Intelligence.
-
2016 - 2020 Rourkela, India
National Institute of Technology, Rourkela
B.Tech, Computer Science and Engineering, GPA 3.7/4.0
- Advised by Dr. Anup Nandy. Undergraduate thesis: Human Gait as a Cognitive State Predictor, in the Intelligent Computing and Computer Vision research group.
- Student Mentor, Institute Counselling Services (2018 to 2020). Member of the institute football team.
-
2003 - 2016 Rourkela, India
St. Paul's School, Rourkela
ISC (Class XII), 96.8%; ICSE (Class X), 97.2%
- Deputy Head Boy (2012 to 2013). School football team, swimming and athletics.
Publications
-
2026 SP-Rank: A Dataset for Ranked Preferences with Secondary Information
KDD 2026, Datasets and Benchmarks Track (Poster)
Hadi Hosseini, Debmalya Mandal, Amrit Puhan. Lead author; authors listed alphabetically.
-
2025 Surprisingly Popular Voting with Concentric Rank-Order Models
The ACM Web Conference (WWW) 2025, Main Track (Poster), pp. 3026–3036
Hadi Hosseini, Debmalya Mandal, Amrit Puhan. Lead author; authors listed alphabetically.
-
2024 The Surprising Effectiveness of SP Voting with Partial Preferences
NeurIPS 2024, Main Track (Poster), pp. 2787–2829
Hadi Hosseini, Debmalya Mandal, Amrit Puhan. Lead author; authors listed alphabetically.
-
2021 Two Novel Methods for Multiple Kinect v2 Sensor Calibration
CVIP 2021, Springer, pp. 403–414
Sumit Hazra, Manasa Pisipati, Amrit Puhan, Anup Nandy, Rafał Scherer. Equal contribution with Manasa Pisipati.
-
2019 A Dynamic Hand Gesture-Based Password Recognition System
ICCIS 2019, Springer, pp. 21–34
Manasa Pisipati, Amrit Puhan, Arun Kumar, Vijay Bhaskar Semwal, Himanshu Agrawal. Equal contribution with Manasa Pisipati.
Academic Service
- Reviewer, The ACM Web Conference (WWW) 2027, 2025 (Main Track)
- Reviewer, Internal Amazon Science Conference 2026
- Reviewer, KDD 2026, Datasets and Benchmarks Track
Honors and Awards
-
2025 Best MS Thesis Award in Artificial Intelligence
Penn State University
Awarded at the third annual Penn State AI Awards, judged by a panel of Penn State AI experts.
-
2024 NeurIPS 2024 Scholar Award
Neural Information Processing Systems
Given to a select group of early-career researchers, with a grant covering registration and travel to attend and present at NeurIPS 2024.
-
2021 Fast Learner Award
SAP Labs India
-
2016 Top 1% in JEE Main
Ministry of Education, Government of India
Among about 1.4 million candidates.
-
2012 National Talent Search Examination Scholarship (All India Rank 245)
NCERT, Government of India
Awarded a national rank of 245 out of more than 1 million candidates. Received throughout middle school, high school and undergraduate studies.
-
2012 Junior Mathematics Olympiad Scholarship (State Rank 8)
Institute of Mathematics and Applications, Government of Odisha
Presented by the Chief Minister of Odisha.