Kanika Gupta

Energy-Aware Machine Learning for Large-Scale Search Systems

Abstract:

The environmental and operational costs of training and serving large-scale AI systems have sparked a critical need for energy-efficient machine learning. This article investigates strategies for reducing energy consumption in large-scale search systems without sacrificing accuracy or responsiveness. To explore a combination of techniques including model quantization, sparsity induction, low-rank adaptation, and intelligent routing within multi-expert models to minimize energy use during both training and inference. The article demonstrates that properly implemented optimization techniques can significantly reduce energy requirements while maintaining search quality. To examine system-level approaches such as adaptive query processing, hardware-software co-optimization, and distributed system considerations that complement model-level optimizations. Furthermore, discussing emerging tools for carbon accounting in ML pipelines and outline a framework for integrating sustainability as a first-class design principle in AI infrastructure. The article suggests that organizations can achieve substantial environmental and economic benefits by implementing energy-aware ML techniques across the search stack without compromising user experience.

Profile:

Kanika Gupta is a seasoned Software Development Engineer II at Amazon with over 7 years of experience building mission-critical systems that protect and scale one of the world's largest e-commerce platforms. Based in Seattle, she specializes in large-scale distributed systems, applied machine learning, and end-to-end architecture design, with deep expertise in Java, AWS services, and data-intensive workflows.
At Amazon, Kanika has been instrumental in advancing Brand Protection initiatives that directly impact customer trust and marketplace integrity. She led the development of an innovative Continuous Proactive Infringement Discovery Framework, creating a multi-modal LLM-based semantic search system that discovers over 500,000 potentially infringing product listings daily across Amazon's global catalog of ~300 million active users. This system achieves remarkable precision by combining textual and image signals through vector embeddings and high-dimensional similarity queries, helping block 99%+ of suspected infringements before brands even report them.
Her technical leadership extends beyond detection systems. Kanika designed the Change Impact Analyzer (CIA), which eliminated manual audit overhead by automating the validation of 700,000+ rule changes, and created a Graduated Enforcement framework that introduced seller-friendly grace periods, resulting in 50%+ of infringing listings becoming compliant within the first month. She also architected Amazon's first Infringement Audit Framework, enabling real-time auditing of over 100,000 ASINs daily with sub-2-second query latency.
Prior to her current role, Kanika gained valuable experience at Infosys Limited, where she customized banking CRM solutions for major clients including ICICI Bank and AXIS Bank. Her academic foundation includes a Master of Science in Computer and Information Science Engineering from the University of Florida (GPA 3.95/4.00) and a Bachelor of Engineering in Computer Science from Punjab University.
Kanika's research background includes cybersecurity work at the Florida Institute for Cybersecurity, where she conducted dynamic security analysis of medical applications. Her combination of theoretical knowledge, practical experience, and innovative problem-solving has made her a key contributor to Amazon's mission of maintaining a trustworthy shopping experience at unprecedented scale.