Uma Maheshwari Earati

AI Application Engineer with 2+ years of experience building production-oriented Generative AI systems using Python and FastAPI. Strong understanding of Machine Learning (ML) and Deep Learning (DL) fundamentals, including transformer architectures and model evaluation techniques. Hands-on experience designing modular backend services integrating Large Language Models (LLMs),Retrieval-Augmented Generation (RAG), and agentic workflows into deployed applications. Focused on scalable architecture, LLM evaluation, CI/CD automation, and reliable AI system design.

• Programming & Backend: Python, FastAPI, REST APIs, Pydantic, API validation, structured error handling Generative AI: LLMs, RAG, LangGraph, LangChain, Prompt & Context Engineering, Tool Calling, Structured Outputs, ML/DL fundamentals
• AI Observability & Evaluation: LangSmith tracing, LLM evaluation workflows, prompt performance monitoring, execution tracing
• Infrastructure & DevOps: Docker, AWS ECS, GitHub Actions (CI/CD), env-based config, containerised deployments
• Data & Retrieval Systems: Embedding pipelines, CLIP, Qdrant, FAISS, semantic search, vector indexing, SQL basics

RAG-Based Document Analysis Platform

• Designed and implemented a modular production-oriented RAG system using FastAPI to deliver LLM-powered document analysis services via REST APIs.
• Built structured retrieval pipelines including document chunking, embedding generation, and FAISS-based similarity search with controlled context injection.
• Integrated structured prompt templates and Pydantic-based output validation to improve reliability, schema enforcement, and controlled generation.
• Implemented LangSmith tracing, along with structured logging and defensive error handling, to improve observability and debugging.

Optimised the system by:

• Reducing token usage through retrieval-grounded responses and controlled context windowing
• Reducing query latency by eliminating RAG reconstruction per request and enabling asynchronous LLM execution
• Improving throughput by reusing LLM and FAISS instances instead of per-request reinitialisation
• Minimising infrastructure overhead through efficient resource management

Containerised using Docker and deployed to AWS ECS with CI/CD automation via GitHub Actions for scalable and consistent deployments.

Multimodal Semantic Image Search System

• Designed and implemented a production-oriented semantic image retrieval system using CLIP embeddings and Qdrant for similarity search.
• Developed modular backend components with structured logging, error handling, and singleton-based client management to ensure maintainability and production readiness.
• Built scalable embedding and indexing pipelines enabling text-to-image and image-to-image retrieval, with structured collection management and batch indexing.
• Integrated a lightweight LLM-based query rewriting layer with intent detection, input length control, and in-memory caching to optimise LLM utilisation.

Optimised the system by:

• Reducing token usage, API calls, and inference latency
• Reducing memory and storage overhead through on-disk vector persistence and minimal metadata design

Developed a Streamlit interface for controlled testing and qualitative evaluation of retrieval performance.

Agentic AI Research Workflow System (LangGraph + FastAPI)

Designed and implemented a production-oriented Multi-Agent AI Orchestration System using LangGraph to manage structured, multi-step LLM workflows (analyst generation → interview → web search → section writing → report synthesis).

Developed modular backend services with FastAPI, ensuring clear separation between Core AI Logic and Production Concerns (structured logging, configuration management, exception handling).

Integrated Tool-Based Web Search (Tavily) using LLM-Driven Structured Query Generation, enabling controlled external data injection and reliable execution.

Implemented Token Usage Tracking (prompt, completion, total tokens) per node to monitor cost and optimise LLM Utilisation.

Optimised the system by:

• Reducing Prompt Size via controlled context injection (top-3 search results)
• Reducing redundant external API calls via in-memory search result caching
• Improving Response Latency using deterministic workflow execution

Automated structured report generation with export to DOCX and PDF formats.

• Unit testing using Pytest (API and workflow validation)
• Structured logging, tracing, and LLM execution monitoring using LangSmith
• API validation and schema enforcement using FastAPI and Pydantic
• Performance optimisation through retrieval grounding and token efficiency strategies

Full UK work authorisation – no sponsorship required