Soumendu Bhattacharjee

Built LangGraph-based multi-agent conversational AI systems for enterprise data Q&A and workflow automation. Developed Snowflake Cortex Agent framework integrating structured, unstructured, and geospatial data for advanced analytics. Designed deep learning audio denoising pipeline to improve multilingual speech-to-text transcription accuracy. Implemented intent classification using embedding models and semantic similarity for intelligent query routing. Built recommender systems using ALS and Two-Tower neural networks; evaluated with Precision@K, NDCG, AUC. Performed end-to-end LLM and RAG evaluation using G-Eval and RAGAS; optimized via prompt engineering, caching, and retrieval augmentation. Deployed scalable AI/ML pipelines on AWS SageMaker with experiment tracking and production monitoring.

Performed entity extraction from PubMed papers for cancer research using LLMs with BioBERT and SNOMED CT ontology-based hybrid approach. Researched and implemented U-Net, Mask R-CNN, and Mask2Former for semantic, instance, and panoptic segmentation of satellite images.

Fine-tuned LLMs for multi-label classification on large healthcare documents in Azure Blob using Spark. Built automated preprocessing and summary generation pipelines with Azure Data Factory and Azure DevOps.

Led LLM + Knowledge Graph integration for enterprise RAG Q&A systems, dramatically improving retrieval quality. Built LangGraph research agents, LoRA fine-tuned LLMs, and graph-based RAG pipelines for complex document Q&A. Developed report generation for large legal documents and news articles using LLMs. Managed Airflow and Cron job orchestration; deployed Streamlit apps using S3, Azure Blob, and SFTP data sources. Worked with MongoDB, Redis, Kafka, and Postgres for storing and streaming real-time content to agents.

Fault detection in PV systems using Autoencoder, Boosting, RFC, and SVM algorithms with Fourier denoising. Supervised long-term GHI forecasting using multivariate LSTM, XGBoost, and CatBoost models. Generated synthetic time series data using Gretel and TimeGAN; converted tabular data to images for CNN classification. Trained and deployed models on AWS SageMaker with S3 and Glue ingestion; stored results in Snowflake and Redshift.

Tabular classification using Random Forest, XGBoost, CatBoost, AdaBoost, SVM with hyperparameter tuning and PCA. Image classification (1500+ classes) using InceptionV3, ResNet50, EfficientNetB7, MobileNet, VGG16/19. Cell and image segmentation using UNet, U2Net, and Stardist; text analysis using BERT, GPT, LSTM, VADER. Built pipelines integrating SQL Server, BigQuery, and DuckDB; leveraged S3, MongoDB, and Redis for storage.

ML-based data analysis and mining of large-scale scientific experiment data; supervised junior researchers.

Published high-impact data analysis results; led ML optimization on large experimental datasets; supervised students.

Data analysis and ML on large-scale physics experiments; Monte Carlo simulations for gamma ray detectors using C++.