About
Builder first.
Researcher always.
I'm Ujjwal Kumar — a Computer Engineering graduate and Software Engineer at JPMorgan, passionate about bridging research and real-world applications in machine learning and data science.
My recent work spans federated learning and privacy-preserving AI, with papers published at ICML and IEEE venues. Outside research, I'm building Tactiq — an autonomous capital allocation engine for Indian equities that backtests regime-aware trading strategies across 150 NSE stocks, with an LLM-based meta-selector in the works to dynamically weight strategies by market conditions.
Outside of work, you'll find me reading, playing guitar, cycling, or travelling — anything that fuels creativity and perspective.
Interests & Skills
Career
Experience & Education
From building financial infrastructure at JPMorgan to shipping hackathon projects — a record of what I've worked on and where I've studied.
JPMorganChase
UNOCUE
CSI – KJSCE Student's Chapter
Education
KJ Somaiya College of Engineering, Vidyavihar
Bachelor of Technology — Computer Science
Research
Publications
My research focuses on federated learning — training machine learning models across distributed devices without centralising raw data.
Federated Learning enables collaborative model training across decentralised data sources — preserving privacy while still benefiting from scale. My work explores algorithms, frameworks, and real-world applications of FL in heterogeneous environments.
View on Google ScholarDMM: Distributed Matrix Mechanism for Differentially-Private Federated Learning using Packed Secret Sharing
Alexander Bienstock, Ujjwal Kumar, Antigoni Polychroniadou
Federated Learning solutions with central Differential Privacy have seen large improvements in utility from the matrix mechanism, while distributed (more private) DP approaches have lagged behind. We introduce the distributed matrix mechanism to achieve the best-of-both-worlds: better privacy of distributed DP and better utility from the matrix mechanism. We accomplish this using a novel cryptographic protocol that securely transfers sensitive values across client committees of different training iterations with constant communication overhead, accommodating the dynamic participation and dropout of users required by FL.
A Comprehensive Analysis of Inference Attacks in Federated Learning
Ujjwal Kumar et al.
This paper presents a comprehensive analysis of inference attacks targeting federated learning systems, examining how the decentralised nature of FL introduces unique privacy vulnerabilities. We systematically evaluate membership inference, attribute inference, and model inversion attacks across diverse FL settings, analysing their effectiveness and the conditions under which they succeed. The study provides insights into designing more robust privacy-preserving federated systems.
MetaFRS: Federated Learning based Cold Start Recommendation System using Meta-Learning
Ujjwal Kumar, Dhairya Ameria, Pragya Gupta
Recommendation systems face two core challenges: user privacy and the cold start problem where insufficient data prevents accurate recommendations for new users or items. We address both by combining Graph Federated Learning — enabling distributed model training without raw data sharing — with meta-learning techniques that allow the system to rapidly adapt to new users with minimal data, delivering personalised recommendations while preserving privacy.
Preserving Privacy in Next Keyword Prediction using Federated Learning
Shruti Tyagi, Aditya Pawar, Ujjwal Kumar, Dhairya Ameria
Next keyword prediction is widely used in mobile keyboards and search systems, but centralised training on user typing data raises significant privacy concerns. We propose a federated learning framework that trains keyword prediction models directly on-device, ensuring raw keystroke data never leaves the user's device. Our approach achieves competitive prediction accuracy while providing formal privacy guarantees, demonstrating that utility and privacy need not be at odds in NLP applications.
Projects
Things I've built
From research frameworks to deployed products — shipped with care, curiosity, and a preference for systems that scale.
Tactiq
A stateful financial decision engine that backtests five regime-aware trading strategies across 150 Indian equities (NSE). Features intelligent universe selection, ATR-based position sizing, and a planned LLM-based meta-selector to dynamically weight strategies by market regime.