Uma Maheshwari Earati

Document Portal – AI-Powered Document Analysis & Chat System

• Built an end-to-end GenAI document portal using Python and FastAPI to upload, analyse, compare, and query documents.
• Automated manual document review by implementing LLM-based document analysis and comparison for PDFs.
• Enabled fast document understanding through a retrieval-based question-answering system over uploaded documents.
• Designed a RAG pipeline with document chunking, embeddings, and FAISS vector search to retrieve only relevant content instead of full documents.
• Reduced unnecessary LLM token usage and cost by grounding responses in retrieved chunks rather than sending entire documents to the model.
• Implemented prompt engineering and structured output parsing to improve response relevance, consistency, and reliability.
• Architected the system modularly , separating API routes, services, prompts, and models for maintainability and interview clarity.
• Containerised and deployed the application to AWS ECS , with CI and CD implemented using GitHub Actions for automated builds and deployments.

LLM-Powered Semantic Image Search System

• Designed and implemented a semantic image search system using CLIP embeddings for text-to-image and image-to-image retrieval , enabling cost-efficient multimodal search without generative inference at query time .
• Built a FastAPI backend to generate and index image embeddings into Qdrant , supporting low-latency similarity search via text or image queries .
• Improved retrieval quality using lightweight LLM-based query rewriting , reducing reliance on expensive generative workflows while maintaining relevance .
• Developed a Streamlit user interface for interactive search and result visualisation.

LangGraph-Based Research Report Generation System

End-to-End Generative AI Application

• Designed and developed an end-to-end Generative AI system using LangGraph to orchestrate multi-step research workflows , enabling controlled execution and avoiding unnecessary LLM calls.
• Implemented an agent-based architecture for automated question generation , web search (Tavily) , expert answer synthesis , and structured report section writing , reducing redundant prompt execution.
• Integrated LLM tool calling , structured outputs (Pydantic models) , and human-in-the-loop execution to control execution flow and limit costly downstream LLM operations .
• Built FastAPI back-end services to manage user sessions , research workflows , and generated reports with predictable request lifecycles.
• Containerised the application using Docker , with a strong focus on modular design , logging , error handling , and clean separation of core logic from production concerns to support maintainability and cost-efficient scaling.

• Joined the organisation as a Trainee and progressed into Team Lead–level responsibilities based on performance and commitment.
• Coordinated end-to-end product listing operations across multiple e-commerce marketplaces, ensuring accuracy and timely execution.
• Conducted periodic website audits, identifying critical content and pricing errors, contributing to a 15% reduction in listing corrections and operational rework costs .
• Managed email newsletter campaigns , including basic design, content preparation, and scheduling.
• Improved inventory accuracy, reducing overselling incidents by 20% and enhancing customer satisfaction.