Udhayan Nateshan Ilangovan

About Udhayan Nateshan Ilangovan

AI Systems & NLP (4+ years): Built agentic solutions end-to-end—from concept to PoC to production—for complex domain-specific use cases. Led cross-functional delivery and mentored junior engineers to scale adoption- Agentic Architecture & Orchestration: Designed multi-agent systems with tool use, planning, retries, budgets, and guardrails using ADK, MCP, and A2A protocols. Delivered full traceability and maintained SLOs across latency, cost, and reliability dimensions- RAG & Knowledge Graph Engineering: Developed GraphRAG pipelines grounded in Neo4j with multi-view NVEmbed embeddings, hierarchical chunking, and custom softmax and coverage-based reranking. Achieved 94% top-1 retrieval and ~70% Level-1 support automation- Transformers & NLP Expertise: Deep understanding of Transformer models, architectures, and internals (e.g, attention, positional encodings, tokenization). Experienced with all major NLP concepts and models across tasks like classification, generation, reasoning, and summarization using frameworks like Hugging Face, spaCy, and NLTK- Fine-Tuning, Evaluation & Infra: Fine-tuned Llama 2/3/3.1/3.2 (≤70B) on vLLM with H100 GPUs using LoRA, RAFT, RLHF, FSDP, and DP. Reduced training time by 50% while enabling low-latency inference. Replaced BLEU/ROUGE with LLM-as-Judge (100% test coverage, 85% reduction in manual QA). Skilled in AWS (Lambda, DynamoDB, SageMaker, AppSync), Kafka/Flink, Python/Java/SQL, Agile/Scrum; contributor to patents and whitepapers.

Experience

1. Ai Engineer

   KLA

   Jun 2024 — Present · Ann Arbor, MI, US

   Led the design and deployment of a domain-specific AI troubleshooting system to automate ~70% of Level 1 support using retrieval-augmented generation (RAG), LLM fine-tuning, and multi-agent orchestration- Developed GraphRAG pipelines by converting thousands of manuals and service logs into structured Neo4j graphs with tool, part, and subsystem traceability- Built an ingestion pipeline that chunked documents by hierarchy (title, section, subsection), generated summaries, and embedded four data views: raw, preprocessed, summarized, and titles using NVEmbed- Engineered a retrieval pipeline with eight specialized agents performing multi-perspective semantic search across all embeddings, combining results via custom softmax and coverage reranking, achieving 94% top-1 retrieval accuracy- Captured expert troubleshooting knowledge by turning human responses into graph facts using a human-in-the-loop ingestion tool for long-term reuse- Fine-tuned LLaMA 70B/90B using LoRA, RAFT, and Chain-of-Thought reasoning, reducing training time by 50% through GPU parallelization (H100) and optimized data pipelines- Created robust datasets through automated question–answer–reasoning generation with multi-view input strategies, improving model generalization and minimizing overfitting- Deployed fine-tuned models into production with custom chunking strategies, token window optimization, and integration into a downstream multi-agent reasoning stack- Designed and orchestrated a full multi-agent system using ADK, A2A, and MCP protocols, with agent routing, tool calling, retries, guardrails, and full trace logging- Developed an LLM-as-Judge evaluation framework to replace BLEU/ROUGE with meaning-aware scoring, achieving 100% test coverage and reducing manual validation by 85%- Collaborated cross-functionally with product, support, and ML teams to scale adoption, deliver product milestones, and mentor interns through project execution- Currently working on RL.

Education

• ESAIP, École Supérieure Angevine en Informatique et Productique

  Master of Science (M.Sc.), Computer Security science

  2017 — 2019

• Stevens Institute of Technology

  Master's degree, Applied Artificial Intelligence

• Anna University Chennai

  Bachelor of Engineering (B.E.), Computer and Information Systems Security/Information Assurance

  2012 — 2016