● Developed the Content Creator Multi-Agent System to create blog posts and interact with clients using Langchain and Langgraph frameworks utilizing the KnowledgeBase with Neo4j, PostgreSQL, Pinecone for the RAG Pipeline, incorporating OpenAI, GPT-4o, Tavily Search tools, and Semantic Search (PgVectors) technology, which improved accuracy by 30%. ● Implemented an end-to-end LLM system pipeline, achieving 100% HIPAA compliance and enhancing healthcare AI assistant reliability using Guardrails AI and LLMOps platform like CircleCI. ● Designed and developed the GAN model for the Bit-stream Generator and optimized, reduced the model size by 4 times and enhanced the training speed and entropy by 3 times. ● Led the development of the Offline Speech Bot Project utilizing Coqui TTS, OpenAI Whisper, and a fine-tuned LLM based on Llama to process and convert real-time human speech into structured data for healthcare and insurance documents with Faiss, resulting in a 35% improvement in the bot interaction speed. ● Deployed a generative AI solution for personalized patient responses using T5, and GPT-Neo on AWS Bedrock and SageMaker, handling over 1 million monthly interactions and raising patient satisfaction scores by 23%. ● Fine-tuned the Healthcare LLM to adapt the BitNet technology and optimized the model size by 2 times rather than before. ● Developed 3+ reliable, multi-modal RAG systems leveraging Agentic AI and conversational AI, improving retrieval accuracy by 30% with HNSW and Neo4j Knowledge graphs. (OpenAI GPT-4o, ChatGPT, Claude, Gemini, Llama, ChromaDB). ● Developed multiple LLM applications using Amazon Bedrock, focusing on model selection and hyperparmeter tuning to optimize performance across diverse use cases, delivering scalable generative AI solutions with end-to-end integration into production environments, achieving a 50% reduction in deployment time and supporting 5 enterprise clients.

● Spearheaded the full cycle of building NLP applications, from ideation and data collection to model design, training, deployment, and iterative optimization, leveraging advanced LLM serving techniques such as continuous batching, loadbalancing, KV caching for latency reduction (reduced latency by 40%), and parameter-efficient fine-tuning methods like LoRA, achieving a 30% improvement in model inference speed. ● Trained and fine-tuned LLMs across multiple architectures(BERT, T5, GPT) for domain-specific tasks, integrating evaluation pipelines with metrics like ROUGE, BERTScore and BLEU, and custom benchmarks, resulting in robust performance improvements for healthcare and customer-facing generative AI solutions. ● Built and deployed a named entity recognition model for finance data using BERT in TensorFlow and spaCy, extracting 1.5 million financial entities annually and reducing error rate in data extraction from 9% to 2%. ● Created an LLM-powered summarization tool for financial news using GPT, NLTK, reducing analysts’ research time from 3 hours to 45 minutes per report and enhancing team productivity. ● Designed a sentiment analysis model for investment insights on financial news datasets with PyTorch and GCP AutoML, achieving 92% accuracy across 500,000 articles. ● Implemented an LDA and BERT-based topic modeling system using Gensim and AWS, categorizing 10 million financial records yearly and improving search relevance by 42%. ● Optimized the model inference using the new transformer architecture design, resulting in improved the gpu usage by 2 times. ● Pioneered multi-modal machine learning applications by fusing LLMs with vision models, leveraging pre-trained embeddings and cross-attention mechanisms to process and generate responses from text, image, and structured data inputs, achieving state-of-the-art results in hybrid AI applications with a 30% improvement in response accuracy across 3 production use cases.

● Optimized data processing pipelines with PySpark and Apache Airflow, reducing computation time by 30% and processing 10%B of data daily with BigQuery integration. ● Designed a fraud detection system using anymaly detection algorithms and NLP techniques, cutting fraudulent transaction by 40% and saving $2M annually for a financial services client. ● Created an OCR pipeline for client onboarding automation using Tesseract, OpenCV, and Keras, reducing manual entry workload from 100,000 to 30,000 hours annually. ● Increased product engagement clicks by 45% for e-commerce clients by developing a recommendation engine with scikit-learn and PyTorch, reaching over 1 million users monthly. ● Built a text classification system for healthcare document categorization using TensorFlow and NLTK, achieving 97% accuracy on 10,000 documents daily and enhancing data management. ● Improved image recognition accuracy to 98% on client products by optimizing a computer vision system with OpenCV and TensorFlow, reducing annual processing errors by 20,000 instances.