Hemanth Konda

Led six MCP servers exposing internal capabilities to LLM agents (e.g., filings retrieval, transcript search, fundamentals, peer comparison, chart data, research archive search). Built the MCP host to broker tool calls, enforce per-analyst rate limits, and emit audit logs for compliance. Designed LangGraph orchestrator that decomposes queries into retrieval, computation, and synthesis steps with deterministic checkpoints. Fine-tuned bge embeddings for retrieval, implemented hybrid BM25 + dense retrieval with reranking, and deployed on EKS with Bedrock Claude inference. Built PII/PII- NSFW scrubbing and robust monitoring across latency, cost, and quality. Partnered with model risk management on SR 11 7 aligned documentation and ran internal code/regression evaluations before production promotions.

Built Lilly’s Gen AI gateway routing prompts to GPT-4 Turbo (Azure OpenAI), Llama 3 on Bedrock, and self-hosted Mistral 7B. Implemented a two-stage adverse event extraction pipeline (BioBERT NER with GPT-4 Turbo verification) and a RAG-based literature review assistant. Deployed Mistral 7B on Bedrock and SageMaker for latency-sensitive tasks, established an evaluation framework (Ragas, TruLens) and gated changes behind passing evaluation runs. Implemented guardrails to block PHI leakage and unsafe medical advice, and created annotation tooling for medical reviewers with active learning loops.

Built the central ML platform on AWS SageMaker, including a feature store, CI/CD pipelines, and monitoring. Implemented drift and performance monitoring with Evidently AI, and built batch/real-time deployments. Migrated legacy SageMaker pipelines to standardized templates using Step Functions and Feature Store. Collaborated with HIPAA security teams to design IAM, VPC endpoints, KMS integration, and audit logging; produced model lifecycle playbooks and governance documentation.