Uz Zama Khamar

Designed and built an end-to-end Retrieval-Augmented Generation (RAG) pipeline, ingesting data from external APIs, performing chunking, transformation, and generating synthetic data to be indexed into the Databricks vector store. Developed scalable vector ingestion jobs using PySpark and implemented Databricks Mosaic AI Vector Search to enable enterprise-grade semantic search and retrieval. Created prompt-tuned instances of CustomGPT and integrated with Databricks vector store endpoints. Collaborated with OpenAI and Databricks support teams to troubleshoot and resolve enterprise integration challenges, including domain authorization mismatches, vector query latency, and performance optimization. Led the transition from traditional Databricks vector-store-based RAG to Model Context Protocol (MCP)-based RAG to address token limitations, improve accuracy and reduce hallucination. Architected MCP tools with Pydantic models and output parsers to guarantee consistency across all the MCP tools domain while mentoring other developers on design and implementation. Developed MCP TestBench, an internal GenAI evaluation tool developed using DeepEval to evaluate the application performance. Integrated MCP TestBench into GitLabs CICD pipeline to evaluate merge requests before merging any code into the application repository. Supported deployment teams during rollout, ensuring smooth releases, secure connectivity between components, and improved overall system reliability. Improved deployment stability by debugging MCP client-server communication issues, optimizing Kubernetes pod allocation, and hardening MCP client configurations against network interruptions. Implemented a multi agent POC called SQL Agent that can dynamically translate natural language query to SQL query, validate, execute and analyze the result using Langgraph. Implemented another multi agent POC called Auth Agent that classifies the query based on users access levels using CrewAI multi agent framework. Owned agile delivery activitiesSprint planning, backlog refinement, demos, code reviewsand collaborated cross-functionally to integrate domain expertise.

Fine-tuned YOLOv8 Image Segmentation on UI elements dataset to detect UI elements on the screen for EZYBot. Worked on image processing algorithms using OpenCV and Google OCR, resulting in a surface automation tool named EZYBot that reduced automation script development time by 80% and eliminated coding completely. Collaborated with cross-functional teams to integrate EZYBot with various automation platforms, enhancing its versatility and adoption across different workflows. Developed an LLM Framework that can breakdown given tasks into subtasks and resolve them by making accurate API calls from given EzyBot OpenAPI specification using LangChain. Developed LLM functionality to automatically fill global parameters needed for making API calls while prompting the user if data is not present. Automatically makes requests to safe API calls while gets user confirmation for sensitive API calls. Added memory functionality to remember users' chat history across multiple sessions. Implemented output parsing to verify the correctness of the received responses. Integrated LangSmith for tracing LLM requests and performance analytics. Added a RAG (Retrieval Augmented Generation) based feature using Croma vector database to resolve queries of Bot Developers from multiple documentations, PDFs, and manuals. Created insightful visualizations of graph data using Gephi, aiding interpretation of complex clusters in the graphs. Executed projects on Hidden Markov Models (HMMs) and Long Short-Term Memory (LSTM) networks for sequence-dependent predictions. Deployed Machine Learning models as endpoints via Amazon SageMaker on AWS, created RESTful microservices using Python Flask.

Collaborated closely with product management to conceive and craft Tableau dashboards tailored for executive-level insights. Collaborate with backend developers to create server-side logic using Java. Optimized data migration between dealerships and OEMs (General Motors), reducing downtime from 3 hours to 35 minutes. Worked closely with frontend developers to implement user interfaces using HTML, CSS, and JavaScript. Applied design patterns to improve code organization, reusability, and maintainability. Developed Java APIs and microservices to handle data requests and responses between the frontend and backend. Executed SQL queries on Oracle to manage and retrieve data efficiently. Used Git version control systems for tracking changes and collaboration.