Utkarsh Kumar

Generative AI (GenAI) and ML Engineer with 4+ years building production agentic AI systems, Retrieval-Augmented Generation (RAG) pipelines, and Large Language Model (LLM) applications across enterprise clients. Hands-on with OpenAI Agents SDK, AutoGen, CrewAI, and LangGraph - function calling, tool orchestration, semantic caching, deep search agents, and trainable persona agents on AWS and Azure.

Open to relocation within India: Pune, Hyderabad, Gurugram.

Data Scientist - Fractal - Bengaluru, India

(2024-12)

• Built an agentic AI chatbot for a healthcare sales client using OpenAI Agents SDK, featuring semantic caching to eliminate redundant LLM calls, a deep search agent mode for complex multi-step queries, and instant RAG with file upload for on-the-fly document ingestion and retrieval.
• Engineered a multi-source persona-based agent system with trainable agents that adapt through user feedback, enabling domain-specific conversation styles across healthcare sales verticals and eliminating manual prompt re-engineering per client.
• Designed and built a multi-agent system using AutoGen (v0.4+) orchestrating Researcher and Writer agents via SelectorGroupChat with Tavily integration, reducing manual research effort by 40% and improving content generation speed by 30%.
• Deployed AI agent workflows using CrewAI with planner and retrieval agents, increasing automation efficiency in chatbot workflows by 30% and enhancing user satisfaction scores by 25%.
• Collaborating with senior data scientists on an Audience Builder ML solution for the Meta team, applying ML-driven segmentation to improve marketing campaign targeting and reduce wasted ad spend.
• Architected a high-throughput data refresh pipeline using Microsoft Graph API on Azure Batch Service, leveraging concurrency and multithreading to reduce refresh time by 70%.

Data Scientist - Accenture - Bengaluru, India

(2021-05 - 2024-12)

• Built a Retrieval-Augmented Generation (RAG) document retrieval system on AWS integrating LangChain, Redis as vector DB, Hugging Face embeddings, and Azure OpenAI for Large Language Model (LLM) inference, significantly improving search accuracy for enterprise ticket resolution.
• Engineered a Spacy NER PII detection layer masking sensitive data in open-text fields before every LLM call, achieving 95% accuracy and eliminating compliance risk across 100% of production user queries.
• Developed an NLP and ML classifier combining TF-IDF and Random Forest to categorize ServiceNow tickets across multiple catalog items with 82% accuracy, reducing support team effort by 30%.

Bachelor of Technology (B.Tech) - Computer Science and Engineering - KIIT University (2017 - 2021)