Kusum Poonia

Innovative Generative AI Engineer with 3 years of experience architecting agentic AI systems and scalable

LLM-powered solutions (RAG, LangChain, VLLM) that drive intelligent automation. Expertise in building multi-agent workflows, advanced AI evaluation frameworks, and secure, production-grade AI architectures,

delivering measurable improvements in efficiency and system performance. Strong foundation in Computer

Vision and deep learning, developing real-time intelligent systems, with hands-on experience in cloud platforms (AWS, Azure, GCP) for end-to-end AI deployment at scale.

Generative AI Engineering & Machine Learning Engineer at XENONSTACK PVT LTD (2023-05 – Present)

Engineered advanced AI systems including multi-agent architectures, LLM evaluation frameworks, and machine learning solutions.

• Engineered a multi-agent NL-to-SQL system where agents handle query understanding, SQL generation, and visualization.
• Built an ITSM multi-agent system for automated ticket lifecycle management (creation, classification, and resolution), achieving an 80% improvement in operational efficiency.
• Built a multi-agent architecture where a VLLM agent extracts vessel data, a validation agent filters relevant information based on predefined fields, and a mapping agent structures and stores the data in databases.
• Developed a multi-agent email processing system automating document validation, TDS checks, and approval workflows, reducing manual effort and turnaround time significantly.
• Designed an evaluation framework for multi-agent systems using Langfuse trace monitoring.
• Developed an agent evaluation framework integrated with CI/CD pipelines, where each deployment triggers automated evaluation of agentic system using single-turn and multi-turn conversations.
• Implemented custom evaluators to assess agent trajectory accuracy, goal completion, and multi-step reasoning, improving reliability and performance of agent workflows.
• Designed a secure agentic AI architecture with an LLM gateway layer enforcing Responsible AI policies, including prompt injection detection, jailbreak prevention, and content safety checks.
• Integrated an agent gateway layer into the workflow to perform pre and post LLM safety checks, including input validation, safety filtering, and output moderation, improving system robustness.
• Developed a YOLOv8-based object detection system for real-time inventory management of alcohol bottles, achieving 97% brand and object detection accuracy. Enhanced low-confidence detections with generative AI, achieving an overall accuracy of 99%.
• Developed a multi-agent computer vision pipeline for real-time surveillance, integrating motion detection, person detection, tracking, and face detection, achieving ~30% improvement in processing speed and system efficiency.
• Designed and implemented an anomaly detection system using Isolation Forest to monitor vessel trajectories, alerting Dubai government officials to unusual activities with 94% accuracy.

Master of Computer Application in Computer Application – Chandigarh University (2021-08 – 2023-07)

Bachelor of Science in Science – Aadarsh Mahila Mahavidyalaya (2018-08 – 2021-07)