At FairAds AI, I dove deep into the world of computer vision and real-time AI systems. My primary focus was building a sophisticated gender and age estimation pipeline using the DETR (Detection Transformer) architecture combined with MiViOLo for demographic analysis. I curated and annotated a custom dataset, then fine-tuned the model using PyTorch, experimenting with various data augmentation techniques and hyperparameter optimization strategies to push the boundaries of accuracy.

Beyond model development, I architected a production-grade event-driven microservices architecture using Apache Kafka for real-time data streaming. The system was designed to handle high-throughput scenarios, with services deployed and orchestrated using Docker and Coolify. I implemented robust monitoring and logging with Prometheus and Grafana, ensuring system reliability and observability across all services.

One of the most exciting challenges was developing a multi-agent AI orchestration system using the Model Context Protocol (MCP) and CrewAI framework. I designed specialized agents for data analysis, report generation, and business intelligence, each with distinct roles and capabilities. The system leveraged LangChain for agent coordination and FastAPI for the backend server, creating an automated pipeline that transformed raw business data into actionable insights.

My journey at AgenixAI was all about pushing the boundaries of enterprise AI and document intelligence. I spearheaded the development of a comprehensive Retrieval-Augmented Generation (RAG) pipeline that revolutionized how the company handled document queries. Using Llama Cloud Parser for advanced document parsing and LangChain for orchestration, I built a system that could understand context and retrieve relevant information with remarkable precision. The pipeline integrated seamlessly with OpenAI's GPT-4 for generation tasks.

I architected an enterprise-grade semantic search engine from the ground up, leveraging Pinecone vector database for efficient similarity search at scale. The system used OpenAI's ada-002 embeddings to transform documents into high-dimensional vectors, implementing sophisticated metadata filtering and hybrid search capabilities. I optimized the indexing strategy and query pipeline using Python and FastAPI, ensuring lightning-fast retrieval even with massive document collections.

Perhaps the most challenging project was building a multimodal document intelligence system that could process diverse document formats. I integrated GPT-4 Vision for visual understanding and Unstructured.io for document preprocessing, creating a unified pipeline that handled PDFs, images, scanned documents, and more. The system employed OCR techniques, layout analysis, and entity extraction using custom spaCy models. I also implemented data validation pipelines with Pydantic and deployed the entire system on AWS using Docker containers.

During my research internship at Lokmanya Tilak College of Engineering, I worked under Prof. Sanjivani Deokar on Conditional Diffusion Models, I focused on advancing techniques for high-fidelity image generation. The project centered on developing innovative approaches to enhance the quality and accuracy of generated images through diffusion-based methods.

A key achievement of this research was the development of a novel attention-based conditioning mechanism, which resulted in an 11% improvement in FID scores compared to existing baseline approaches. To optimize the training process, I implemented sophisticated techniques including mixed-precision training and gradient accumulation, which significantly enhanced the overall computational efficiency of the model training pipeline.