Manali Maniyar

About Manali Maniyar

SDE I @ Amazon Prime Video | ML Infrastructure & PersonalizationI build and scale machine learning systems that power the personalized recommendations millions of customers see every day on Prime Video. I’m driven by the challenge of turning massive behavioral data into real time, meaningful experiences and designing the infrastructure that makes those experiences reliable, fast and global.At Amazon, I\'ve contributed to evolving Prime Video’s personalization stack by working on foundation model powered pipelines and a suite of lightweight, fine-tuned models for behavioral prediction. These systems operate on large scale datasets and serve low latency inferences across regions using AWS native tooling.Some areas I\'ve worked on:• ML pipeline orchestration using Metaflow, Step Functions and SageMaker• Real-time inference workflows with Lambda, ECS Fargate and SageMaker Endpoints• Feature and embedding generation systems at nearline and offline scale• Observability tooling for performance, data quality and latency across environmentsI\'m particularly interested in:Scalable ML systems • Model distillation & fine-tuning • Low latency inference • ML observability • Foundation models in personalization • Cloud-native MLOps • AWS Infrastructure • Applied Generative AIAlways open to connecting with engineers, researchers and builders working on large scale ML systems and personalization technologies.

Experience

1. Software Engineer i

   Amazon

   Sep 2024 — Present · Seattle, WA, US

   Built and scaled real-time ML inference services using AWS SageMaker, ECS Fargate, Lambda and Kinesis, delivering low-latency predictions (~<120ms p95) for personalized video recommendations, impacting 100M+ Prime Video users globally- Led infrastructure automation for multi-region deployment using AWS CDK and custom constructs, provisioning load balancers and regional DNS entries, supporting services in 3 availability zones for fault tolerance- Enhanced pipelines using Apache Spark, Java, and Scala, processing 2B+ watch and purchase events/day to generate real-time user/item embeddings for personalization models- Instrumented full observability into ML systems (CloudWatch, custom metrics), improving SageMaker error detection latency by 40%, reducing inference timeout failures by 25%, and enabling proactive alerting on stale model deployments.

Education

• University of Southern California

  Master's degree, Computer Science

• Dwarkadas J. Sanghvi College of Engineering

  Bachelor of Engineering - BE, Computer Engineering

  2018 — 2022