Prajit Muppidi

About Prajit Muppidi

Applied Science Leader who has built and scaled a science team to 12 members, fostering 5 promotions to senior roles. I specialize in translating novel research into business-critical products, having led the launch of Alexa\'s first large-scale production Generative AI system for Fire TV.My work bridges foundational research (including a Best Paper Award-winning system at EMNLP 2023) with robust engineering. I\'ve invented frameworks that use LLMs to optimize their own instructions, improving task accuracy from 62% to 84%, and pioneered constrained decoding systems to achieve 100% reliability for business-critical transactions like subscription management.Core Expertise: Generative AI System Design, Large Language Models (LLMs), Prompt Engineering & Optimization, Multi-Armed Bandits, Dual-Encoder Architectures, and leading high-performing science teams.

Experience

1. Applied Science Manager

   Amazon

   Oct 2022 — Present · Seattle, WA, US

   Recruited, built, and led a high-performing team of 12 applied scientists responsible for delivering foundational Generative AI capabilities for Alexa\'s entertainment experiences. I defined the team\'s technical vision, mentored scientists to achieve 5 promotions to senior/next-level roles, and owned the end-to-end delivery of large-scale AI systems from research to productionGenerative AI & LLM Product Leadership- Led a cross-functional initiative of 25+ scientists and engineers to launch the Conversational Content Discovery (CCD) LLM on Fire TV, Alexa\'s first successful production LLM feature, achieving 95% accuracy on complex, open-ended conversational queries- Spearheaded the technical strategy for the Amazon Music Amp launch, guiding a partner ASR team to improve brand name recognition from 76% to 93% for the live radio app.Technical Innovation & Research- Addressed fundamental flaws in voice-assistant entity resolution by designing a novel dual-encoder retrieval system (EREN) that learns from user listening sessions. This approach achieved a 91% improvement in recall@1 over previous baselines. The underlying research for this system,\"Personalized Dense Retrieval on Global Index,\" was awarded Best Paper at the EMNLP 2023 industry track- Invented a framework (AIR) that uses LLMs to automatically optimize their own instructions, improving API chaining accuracy from 62% to 84%. This platform is now used by over 90 stakeholders across Amazon- Pioneered a Constrained Decoding (CD) system that forces LLMs to generate syntactically correct and business-rule compliant outputs, achieving 100% accuracy on contractual use cases like subscription management- Replaced a conservative, batch-based model update process with a real-time \"test and learn\" system using a multi-armed bandit framework (ERWIN), enabling a 10x faster deployment rate of fixes while maintaining quality

Education

• Coursera

  Deep Learning Specialization, Artificial Intelligence

  2017 — 2018

• Indian Institute of Technology, Kharagpur

  B.Tech+M.Tech (Dual Degree), Mining Engineering

  2010 — 2015

• Columbia University

  Master’s Degree, Operations Research

  2015 — 2017

Skills

• Html
• Statistics
• R
• Sql
• C
• Programming
• Microsoft Excel
• Powerpoint
• Microsoft Office
• Autocad
• Teamwork
• Vba
• C++
• Machine Learning
• Slam
• Matlab
• Windows
• Data Mining
• Python
• Solidworks
• Algorithms
• Java
• Microsoft Word
• Data Science