Nikhil Deekonda

Developed an AI-driven patient engagement platform that empowers clinicians to securely communicate with patients, analyze health data, and deliver personalized care recommendations-integrating AWS Bedrock, LangChain, Redis, and MongoDB for scalable, privacy-compliant deployment. Architected a Mixture of Experts (MoE) architecture on top of a DeepSeekR1-distilled LLaMA 8B model, introducing domain-specific experts fine-tuned with RAG-based datasets, improving contextual accuracy by approximately 30% in internal evaluations. Engineered a high-performance RAG pipeline using HNSW indexing for dense retrieval and cross-encoder reranking for contextual relevance, achieving 40% faster retrieval speed, improved factual grounding across clinical domains. Designed and currently expanding a Neo4j knowledge graph to enable structured reasoning and relationship-based querying, enhancing clinical recommendation precision and supporting longitudinal patient data insights. Fine-tuned large language models for the patient-side conversational interface using synthetic data generated via few-shot prompting with GPT-4o and DeepSeek, ensuring context-aware, guideline-aligned multi-turn interactions that reduced error rates by 20%. Built and deployed backend microservices using FastAPI, containerized with Docker, and orchestrated on AWS EC2-integrating AWS Cognito for secure authentication and Redis caching for low-latency session management.

Developed guardrails, data pipelines, and validation flows for ML/DL models including LLMs; fine-tuned models like GPT, Falcon, LLaMA-2, and Mistral for real-world use cases (PII removal, tone classification, toxic word detection) using LoRA and QLoRA techniques. Built a scalable platform to fine-tune LLMs with multiple techniques and implemented output validation for LLM agents utilizing frameworks such as AutoGen and LangGraph. Designed and executed complex Retrieval-Augmented Generation (RAG) workflows integrating LLMs with vector databases; developed hallucination detection systems leveraging Small Language Models (SLMs) and BERTbased models. Trained a proprietary 350M parameter language model from scratch and deployed fine-tuning, RAG, and validation workflows on AWS using services like EC2, S3, SageMaker, and Bedrock

Employed various Machine Learning models to predict surface distresses that result in Rutting of Pavements

Worked on mapping of Liquefaction hazard sites across India and predicted areas prone to Liquefaction.

Sacks Impact Scholars Graduate Fellowship Holder