Adithi Kallem

About Adithi Kallem

I\'ve always been drawn to the question: what is this data actually telling us?Over the past 5+ years, that curiosity has taken me from building fraud detection systems at PhonePe — protecting millions of users across India — to designing Generative AI and forecasting solutions at Amazon that run across 12 global regions and handle millions of events every single day.I don\'t just build models. I obsess over whether they actually work in the real world — whether they\'re fair, explainable, and reliable enough to stake a business decision on. That\'s led me into everything from RAG pipelines and LLM-powered assistants to capacity planning forecasts and responsible AI frameworks.Some things I\'ve worked on that I\'m proud of:A RAG system that improved AWS documentation search accuracy by 34% — meaning fewer frustrated engineers, more self-service winsFraud detection models that prevented $2.3M in losses while analyzing 50K+ accountsConversational AI handling 150K+ monthly queries with a 78% resolution rateForecasting models with 8.9% MAPE, helping Amazon plan cloud resources across the globeOutside the numbers, I care deeply about building AI that\'s trustworthy — I\'ve led work on bias detection, fairness metrics, and governance pipelines because I believe the \"responsible\" part of AI isn\'t optional.I work primarily in Python, SQL, PyTorch, and AWS, and I\'m always up for a good conversation about ML systems, GenAI, or the messiness of real-world data.Let\'s connect — whether you\'re a fellow data nerd, building something interesting, or just want to talk shop.

Experience

1. Applied Scientist

   Amazon

   Feb 2024 — Present · WA, US

   One of the projects I\'m most proud of is building a RAG-based search system that made AWS documentation genuinely useful engineers could find answers on their own 34% more accurately, and support tickets dropped by 22%. It sounds like a technical win, but really it was about saving people time and frustration.On the fraud and cost side, I built ML models that sifted through 5 million daily events across 50K+ accounts catching $2.3M in fraud and uncovering $8.5M in cost savings. The scale still amazes me when I think about it.I also led the design of a Responsible AI framework bias detection, fairness metrics, automated evaluation — because I believe AI that you can\'t trust isn\'t really useful AI. That work reduced bias incidents by 41%, and more importantly, it gave the teams around me more confidence in what we were shipping.Forecasting was another big piece I built time-series models using ARIMA, Prophet, and LSTM to help Amazon plan compute capacity across 12 global regions, hitting 8.9% MAPE. Getting cloud infrastructure right at that scale has real consequences, and it was deeply satisfying to get it right.I also built and deployed Conversational AI apps using fine-tuned LLMs that handled 150K+ queries a month with a 78% resolution rate and a 19% lift in customer satisfaction. Seeing people actually get value from something you built never gets old.And behind all of it solid MLOps. Automated monitoring, retraining pipelines, A/B testing via SageMaker the kind of infrastructure that cut model drift by 53% and let the team experiment 38% faster. Good science needs good systems.

Education

• George Mason University

  Master's degree, Data Analytics Engineering

• Sridevi womens engineering college

  Btech, CSE