Ripan Purkait

Productionized computer vision models using ONNX, TensorRT, OpenVINO, Triton, and vLLM with AWQ/GPTQ quantization, reducing inference latency and cloud cost. Built end-to-end MLOps pipelines for vision models including data ingestion, training, CI/CD, model registry, and real-time monitoring for regulated healthcare AI systems. Designed generative and agentic AI pipelines using vision-language models and LangGraph, enabling automated radiology reporting and multi-step clinical reasoning. Implemented real-time inference, agent tracing, and observability using WebSockets, ActiveMQ, and Langfuse for event-driven medical AI workflows. Worked with sensitive EHR and imaging data under GDPR and healthcare compliance guidelines, implementing data anonymization, access controls, audit logging, and secure model serving.

Built autonomous agent-based AI systems using LangGraph, CrewAI, and CopilotKit for customer-facing and internal enterprise workflows. Implemented Self-Recursive, Adaptive, and Agentic RAG architectures to improve contextual retrieval and reduce response latency by 25%. Designed scalable vector and graph-based retrieval pipelines using Neo4j and Azure Cosmos DB to support high-throughput LLM applications. Integrated multi-agent orchestration frameworks (CrewAI, Phidata) to automate business processes and reduce manual intervention.

Built a multimodal RAG-powered healthcare chatbot combining text and vision, improving answer relevance by 35%. Fine-tuned Mistral-7B and Phi-3 vision-language models using QLoRA, DPO, and PPO to optimize accuracy, cost, and latency. Developed LLM-driven web-scraping and data-curation pipelines to create high-quality training datasets at scale. Developed and trained YOLOv5 and YOLOv8 models for real-time object detection and tracking, optimized for live CCTV feeds via RTSP streams using ONNX, TensorRT and deployed model into AI server throiugh docker and kubernetes. Deployed optimized models on edge devices, achieving >95% detection accuracy through hyperparameter tuning, custom post-processing, and model optimization, improving surveillance efficiency by 30%.