Udaya BhaskaraRaju Patsamatla

Architected and refined RAG pipelines and multi-agent frameworks for ingesting Android bug reports and Logcat files into proprietary Google LLMs, automating root-cause analysis at scale. Engineered complex multi-step prompt templates with continuous human-in-the-loop testing, improving semantic accuracy of AI-generated bug summaries and synthetic scam datasets. Managed ingestion & vectorization of Android OS documentation in vector databases for P1-P4 root-cause mapping. Validated AI outputs with Android owners and researchers, refining model logic under SLA targets. Monitored production performance of routing engines and classification agents; debugged MLOps pipeline bottlenecks to ensure continuous delivery and high reliability.

Designed and deployed enterprise-grade Generative AI and agentic solutions using LangChain, LangGraph, AWS Bedrock, and AWS SageMaker, automating complex data retrieval and workflow orchestration. Built RAG LLM applications using Bedrock and OpenSearch for semantic search over enterprise documentation. Integrated MCP servers with Google Drive, Confluence, Jira, building LangGraph agents for enterprise solutions. Optimized SageMaker Multi-Model Endpoints on GPU instances, increasing GPU utilization by ~80% and reducing hosting costs by 66–74%. Built and deployed end-to-end ML pipelines on AWS SageMaker Pipelines, including training, evaluation, Model Registry versioning, and automated deployment with CI/CD integration. Streamlined large-scale data collection and preprocessing using automated Python pipelines, reducing manual effort and improving downstream LLM training quality. Developed advanced prompts, multimodal LLMs, and automation frameworks for training and evaluation. Implemented SageMaker Multi-Model Endpoints to co-host multiple LLM services, raising GPU/CPU utilization by ~80% and reducing inference costs by up to 75%. Leveraged SageMaker inference components and auto-scaling to dynamically adjust model load, reducing idle compute costs by ~50%. Built efficient NLP pipelines with transformer embeddings, NER, and semantic similarity for enterprise use cases.

Fine-tuned LLMs to classify global telephony traffic as scam, spam, or legitimate across multiple languages. Built a multi-agent simulation with Caller, User, Mobile Agent personas to generate multilingual synthetic datasets. Developed dynamic prompt optimization pipelines to improve contextual accuracy, linguistic diversity, and realism of generated data. Designed evaluation workflows to test models against adversarial inputs and concept drift, ensuring high accuracy. Created a scalable AI-generated training pipeline foundational for global scam prevention systems. Built PostgreSQL connectors for automated schema introspection and column description generation, enabling robust RAG workflows. Integrated LangChain with PgVector for semantic schema search, improving SQL generation accuracy. Developed robust dynamic SQL generation and validation logic based on user intent and metadata. Benchmarked GPT-4 variants and Gemini models using the BIRD dataset for accuracy, latency, and cost. Created automated post-query analytics pipelines delivering actionable business insights to stakeholders. Integrated MCP servers with Google Drive, Confluence, and Jira for secure enterprise document access workflows. Developed LangGraph agents with stateful conversation logic for context-aware data retrieval systems. Combined semantic search with tool invocation to support real-time enterprise workflow automation. Built validation and monitoring pipelines to ensure connectivity, accuracy, and response reliability checks. Enhanced agent performance iteratively through stakeholder feedback and prompt optimization cycles. Extracted textual and visual content using CLIP, LLaVA, and GPT-4V, merging multimodal data into unified retrieval documents. Engineered adaptive chunking and storage strategies in PostgreSQL with PGVector to improve retrieval precision. Built LangGraph agents to orchestrate OCR, retrieval, generation workflows dynamically based on query context. Developed an end-to-end multimodal RAG pipeline for context-aware document question answering. Optimized inference pipelines for low-latency responses across large document datasets. Created COBOLCodeBench, a benchmark evaluating 15+ LLMs on COBOL modernization tasks. Designed a two-phase data pipeline: real-world COBOL corpus analysis and expert-led ground truth creation for 46 programming tasks. Conducted ablation studies across prompt formats, identifying “Instruction with Context” as the most effective. Built an automated evaluation framework using CSR, ESR, Results Match, BERTScore for holistic LLM assessment. Published research on LLM-based legacy system modernization and code quality evaluation.

Infosys Certified AWS Solution Architect

Infosys

Infosys Certified Generative AI Master

Infosys

Microsoft − Develop Generative AI Solutions with Azure OpenAI Service

Microsoft

Google Cloud Certified − Generative AI Leader

Google Cloud