Biswas Nandamuri

Architected and developed autonomous AI agent systems for medical document intelligence using LangGraph and Model Context Protocol (MCP) with Vertex AI Gemini. Built stateful, human-in-the-loop workflows with guardrailed production deployment patterns. Engineered event-driven medical document pipelines on Azure Kubernetes Service using Service Bus-triggered KEDA ScaledJobs and Dockerized Go/Python services for OCR, billing, medical events, chat, indexing, and drafting. Designed production-grade medical RAG pipelines using LanceDB with dense embeddings, metadata filtering, hybrid retrieval concepts, re-ranking, and vision enrichment to improve retrieval quality from OCR and medical artifacts. Led LLM/ML evaluation strategy using MLflow and Langfuse datasets (golden test sets, hold-out splits, evaluator scoring) with CI/CD release gates to benchmark recall, accuracy, latency, and cost. Improved medical event extraction recall from 77.1% to 82.5% and reduced missed medication recognition

Shipped LLM-powered content generation features that reduced authoring time from hours to minutes using FastAPI and LangChain with map-reduce summarization and DSPy prompt optimization plus automated quality feedback loops. Designed and deployed BI solutions including 15+ Tableau dashboards and an SQL-based analytics platform serving 5K+ daily users, replacing a legacy 10-minute aggregation system using PostgreSQL materialized views and BigQuery optimizations. Built automated reporting with Advanced Excel (Power Query, Pivot Tables, formulas) eliminating ~15 hours of manual work weekly and reducing entry errors by 90%. Implemented observability and evaluation hooks into LangGraph agents with latency tracking, validation, and auto-retries, reducing silent agent failures by 65%. Reduced LLM API costs by half and increased throughput 4x using async FastAPI, Redis caching, and intelligent routing between GPT-3.5/GPT-4/Claude models. Optimized RAG latency using multi-stage retrieval (FAIS

Co-led SMART (Social Media Analytics and Reporting Tool) funded by NATO to analyze correlations between IoT sensor data and social media sentiment for real-time situational awareness. Built a FastAPI + React backend for interactive visualization using Tableau-style dashboards. Developed PySpark/Databricks pipelines to ingest and process text and sensor data with BigQuery and GCP Cloud Storage. Applied transformer models (BERT, BART, T5) for topic detection and summarization, generating concise daily insights. Implemented SpaCy-based entity redaction for PII removal and GDPR compliance. Used MLflow with Docker-based experiment tracking for reproducibility and dataset versioning, and contributed to technical documentation on multimodal AI for environmental event detection.