Working on production-scale LLM systems, RAG pipelines, and enterprise AI platforms with a focus on reliability, scalability, and evaluation.

●Designed and deployed RAG pipelines integrating vector databases and LLM APIs, improving retrieval accuracy by 32% and reducing hallucinations by 27%

●Fine-tuned transformer-based LLMs using domain-specific datasets and RLHF, increasing response relevance by 35%

●Built automated LLM evaluation frameworks measuring hallucination, latency, and answer quality, improving reliability by 30%

●Developed scalable AI microservices using Python, FastAPI, Docker, and Kubernetes for production inference systems

●Engineered RAG architectures using LangChain, LlamaIndex, Pinecone, and FAISS for semantic search

Implemented distributed ML pipelines using PyTorch, Hugging Face, and Ray Serve for GPU-based inference

●Designed monitoring systems using Prometheus and Grafana for model observability and performance tracking

Built autonomous AI agents and multimodal systems for web interaction and task execution using reinforcement learning and LLMs.

●Designed autonomous AI agents for multi-step web tasks, improving task completion accuracy by 34%

●Developed multimodal pipelines combining vision transformers and LLMs for UI understanding

●Implemented hierarchical planning systems converting natural language into structured actions, reducing failures by 27%

●Applied reinforcement learning (PPO, actor-critic) to optimize agent behavior in dynamic environments

●Built scalable data pipelines for large-scale web interaction datasets

●Optimized distributed training and inference on GCP with TPU infrastructure

●Deployed models using Vertex AI with monitoring and automated retraining pipelines

Developed NLP and computer vision pipelines for document processing, automation, and fraud detection in financial systems.

●Built OCR + NLP pipelines extracting structured data from financial documents, reducing manual effort by 70%

●Developed document classification and entity extraction models improving accuracy by 35%

●Designed fraud detection validation rules enabling real-time anomaly detection

●Engineered ML pipelines using TensorFlow, PyTorch, OpenCV, and spaCy

●Deployed containerized ML services using Docker, FastAPI, Kubernetes, and AWS

●Implemented end-to-end MLOps workflows, including CI/CD and model monitoring