UTKARSH TIWARI

Built and deployed production RAG pipelines using LangChain, enabling enterprise document retrieval, context-aware reasoning, and tool-assisted question answering. Fine-tuned open-source LLMs (Qwen Family) using PEFT with LoRA and QLoRA, enabling efficient on-premise inference while significantly reducing GPU memory requirements. Designed and implemented a production multi-agent AI platform using Google ADK and LangGraph with stateful orchestration, conditional routing, tool calling, and memory for customer support and enterprise documentation workflows. Implemented systematic LLM evaluation framework tracking correctness, latency, and hallucination control metrics across agentic workflows, enabling rapid iteration cycles. Improved model accuracy through spatial and temporal analytics, identifying performance gaps and implementing post-processing fixes. Built end-to-end MLOps pipelines for multiple models, handling training, inference, and deployment in production. Tuned CUDA settings for YOLO and MTO models, improving inference speed and GPU efficiency. Developed reusable Python SDKs and internal libraries to standardize model training, inference, and deployment across AI teams. Restructured inference workflow to run projects in parallel, cutting daily processing time by 5 hours. Set up CI/CD pipelines with Bitbucket and Jenkins to automate builds, tests, and deployments. Led migration from service-based setup to cloud-native architecture using Helm and Kubernetes, improving scalability. Consolidated all GPU machines into a single Kubernetes cluster, removing manual GPU management and improving resource usage. Managed GCP infrastructure (GKE, GCS, networking), optimizing cloud operations and reducing annual costs by 10-12%.

Led AI/ML initiatives from design through deployment, shaping product direction and scalability. Designed predictive models using LSTM for AI-driven insights and better forecasting. Designed and deployed a production RAG platform with retrieval optimization, prompt engineering, and evaluation workflows, improving AI response quality and enabling rapid experimentation. Prepared high-quality datasets for LLM fine-tuning and evaluation, leveraging PEFT techniques to improve task-specific accuracy while minimizing inference cost. Built AI/ML features that improved analytics and workflow efficiency, helping customers process data faster and unlock 60% more market opportunities. Debugged performance issues in a core product, boosting cloud performance by 20%.

Built a custom web scraping library that outperformed Scrapy in speed and efficiency. Created an LLM-powered automation platform using RAG to streamline knowledge extraction and reduce manual effort by 30%. Designed and maintained distributed data pipelines processing over 3 million product documents daily with high availability and low-latency processing. Refactored legacy code with OOP principles and design patterns, cutting code duplication by 25% and making the system easier to maintain. Optimized data workflows with Airflow for better task scheduling and parallel processing.

Developed evaluation frameworks and quality metrics for LLM applications, tracking accuracy, latency, and failure cases across production deployments. Built internal tool using open-source network protocol analyzers, contributing improvements back to the community. Created a Kubernetes Job launcher to automatically manage cloud workloads, increasing performance by 40%. Set up tools to improve RESTful API design, documentation, and testing workflows. Implemented secure storage for passwords and keys in the database, improving security score by 32%.