Sourabh Kumar

SOURABH KUMAR LinkedIn |- |-EDUCATION IIT (BHU), Varanasi Chemical Engineering CGPA- 8.89 Varanasi 24/6/2020 SKILLS  LLMs, Multi-Agents, LangChain, MLOPs, AI, ML, DL, Python, Kubernetes, Docker, QLora, PEFT, Jenkins, SQL, MongoDB, Flask, Azure, TensorFlow, PyTorch Serving, Multimodal, System design, AutoGen, LangGraph, OCR, LayoutLMV3, Llama, Quantization, LLM Inference, LLMOps WORK EXPERIENCE RSYSTEMS Data Scientist Landmark Chatbot and Landmark NER    Noida May 2024-Present Developed a Llama3.2 3b fine tuned model to predict the ner tags using 4 bit quantized QLora method using unsloth framework and pushed model to huggingface(link- Sourabh1172 (Sourabh Kumar)) and also finetuned LayoutLMV3 for document classification(also pushed on same huggingface link) Developed a robust chatbot utilizing Multiagents: RAG fine tuned on Llama 3 using QLora(PEFT)and Unsloth for unstructured data, Langchain create_sql_agent for structured queries, and a Multimodal (Llama 3.2) agent for visual data interactions, including graphs, plots, and images. Deployed the entire solution efficiently using Docker and Kubernetes, ensuring scalability and seamless integration. MLOps using Kubeflow    Built an end-to-end MLOps pipeline for CI/CD, experiment management, and comprehensive data/model drift detection for multiple models ( Random Forest and Time Series). Used MinIO for managing versioned data and model storage, and configured Kubeflow Pipelines for preprocessing, training, and model evaluation in conjunction with Kubernetes. Deployed models with Seldon Core for real-time inference, automated workflows with GitHub Actions, and established monitoring with Prometheus and Grafana, alongside Alibi Detect for data drift visualization. The setup was fully implemented on an on-premises Ubuntu server. OPTUM Data Scientist ContractZoom    Noida Sep 2021- May 2024 Semantic Search: Developed an efficient search system using ElasticSearch with Sentence Transformer embeddings, employing techniques like vector store, sharding, and indexing for parallel processing. Enhanced results by re-ranking with Cross-Encoders, followed by clustering with HDBSCAN and topic modeling. Digitization: Extracted entities from documents using NER models based on QLora(PEFT) finetuned Llama storing known entities in CosmosDB. For unidentified entities, employed LangChain's RAG pipeline powered by LLMs. Document Segmentation: Segmented document components (headers, footers, paragraphs, subheadings, tables) using Vision Transformer for image vectors and Grid Transformer for textual vectors. Combined these with OCR bbox coordinates for classification using Cascade R-CNN. Network Pharmacy Digitization    Led extraction of structured data from pharmacy contracts using YOLO and OCR. Designed a table classification model with Hugging Face's BERT to identify network tables. Utilized Llama2 and NER for additional entity extraction with ReAct-based prompts and integrated LangChain's vector store and sequential chain for optimal Llama2 model performance. Deployed solutions via Flask, Gunicorn, PyTorch Serve, Docker, Kubernetes, and Loadbalancer, implementing a CI/CD pipeline with Jenkins Claim Forecasting  Employed diverse time series models (ARIMA, SARIMA, Exponential Smoothing, LSTM) on 5-year historical claims data for future claims prediction. XPRESSBEES Jr. Data Scientist Mar 2021-Aug 2021  Worked on weekly reports on performance of various zones using SQL & Python and completed ML, DL training COURSES  Software Design and Architecture– DeeplearningAI  Azure Solutions Architect Expert– Microsoft  Generative AI with LLMs- Coursera  Machine learning for Production (MLOPS)- Coursera