● Built a Real-Time Data Ingestion Pipeline using Kinesis Data Streams, Glue Streaming, and

Apache Spark Structured Streaming, reducing Quicksight dashboard latency from daily to

sub-minute updates and enabling real-time analytics for stakeholders

● Took initiative to incorporate GenAI into customer analysis workstreams by developing a new

LLM-based service using Apache Spark and AWS Bedrock (Claude 4.5), reducing manual product

analysis effort by 70%

● Engineered a new personalization pipeline for FireTV including training data preparation, model

training and tuning, and experiment analysis leveraging AWS Personalize, increasing

click-through-rate of selected playlists by 15%

● Implemented a regionalized telemetry service called by FireTV devices using API Gateway and

Lambda with token-based authentication, handling 10,000+ of TPS with sub-100 ms latency,

enabling FireTV Channels internationalization to global markets

● Converted performance-intensive offline SQL join of FireTV item, session, and engagement

events into online real-time data hydration pipeline, enabling real-time features such as video

Continue Watching and reducing peak offline Redshift computational load by 30%

● Re-architected Anomaly Detection service to decouple data retrieval from Redshift, enabling

integration with Athena, Spark, and other data sources, expanding adoption across 5+ teams and

reducing incident response time by up to 80%

Designed and implemented an asynchronous AWS Redshift query scheduler for retrieving

information from Amazon Alexa’s Advertising and Engagement Metrics database

Deployed AWS constructs such as Lambda, SQS, Eventbridge, SNS, and DynamoDB using

the AWS CDK and Typescript

Improved precision of scheduled queries from one hour to one minute, and implemented a review

process to increase reliability and safety of new queries

Interned under Dr. Dana Dachman-Soled and graduate student Aria Shahverdi

Used Python, Anaconda, and Jupyter Notebook to analyze racial and gender biases in the NYPD Stop and Frisk and the US Census Adult Income datasets; the Pandas library was used for data processing and the Seaborn library was used for data visualization

Used the Scikit-Learn machine learning library to implement logistic regression and MLP algorithms that attempted to correct biases