Marthala Pavan Kumar Reddy

Leveraging AWS Bedrock, designed and implemented an innovative code conversion pipeline that transforms ODI packages into equivalent PySpark code with over 95% accuracy. Additionally, developed an end-to-end video summarization system for sports content, utilizing AWS services to capture key moments from full-length videos. This system efficiently creates highlight versions under 20 minutes in duration, with the entire pipeline conversion process completed in less than 10 minutes. The solution optimizes cloud resources, ensuring both processing efficiency and cost-effectiveness, while demonstrating the ability to handle large volumes of video content at scale.

Worked on the development of tailored product recommendation systems for a Retail client, excelling in tasks such as suggesting new items and classifying reorder likelihood across 30 sites with a 20M-record datasets for most of the sites. Leveraged WALS for recommendations, AutoML for classification, and seamlessly integrated MLops for efficient deployment with precision 70% and recall 30%(for classification). Automated end-to-end processes using Airflow DAGs.

Designed and implemented a solution for applications where using production data for testing poses security concerns. Specifically, focused on an LLM-based application that generates synthetic data for Tabular data mirroring the characteristics of the original dataset. By utilizing a few samples of authentic data, the application produces synthetic data with similar traits using LangChain and Amazon Bedrock Claude V2 API. Additionally, established a validation pipeline incorporating exploratory data analysis (EDA) and statistical analysis to ensure the generated data faithfully replicates all characteristics of the original dataset.

Developed a Weather Voice Assistant utilizing AWS services, including Lambda, Comprehend, Lex V2, and Redshift DB. Proficiently managed 10 diverse weather intents, ensuring rapid responses within 300 ms and achieving an Automatic Speech Recognition (ASR) accuracy of 93 %.

Created a Document Parsing application utilizing LayoutLM V4 for extracting necessary fields and signatures from 25 different document types. Employed ResNet50 for checkbox detection, and Detectron for stamp identification, integrated seamlessly with AWS services (Lambda, DynamoDB, SNS, S3). Achieved successful deployment of containerized models, enabling user-friendly interfaces upon image submission with the accuracy of 91 %.

Implemented an efficient system for object detection and text extraction in real-time YouTube live news channels, enabling rapid coverage of breaking news ahead of competitors. Utilized YOLOv5 model for object detection and OCR for text extraction, achieving a commendable accuracy rate of 89 %. Deployed and optimized the models for seamless and effective integration.

Developed and deployed a customized loan analysis dashboard with an integrated data pipeline, leveraging open-source SBA datasets and client-specific data sources. This comprehensive solution enables clients to compare multiple banks and analyze custom characteristics, facilitating informed investment decisions and maximizing profitability. The dashboard's user-friendly interface and powerful analytics capabilities provide clients with real-time insights, reducing decision-making time by up to 40%. Additionally, the system's ability to incorporate diverse data sources has improved the accuracy of loan risk assessments by 25%, leading to a 15% increase in successful loan applications for clients. This tailored solution not only streamlines the loan analysis process but also empowers clients to identify optimal investment opportunities, resulting in an average 20% boost in their investment returns.