LLM-RAG (text, image, multimodal), E2E GenAI

Tech lead: end-to-end machine learning for document understanding

1. Initiated and led a project on developing deep-learning based hybrid models for high fidelity physics-based porous-media flow simulations: - Developed deep-learning based proxy models to replace constitutive equations in coupled PDEs using TensorFlow/Keras functional APIs- Performed distributed training on multi-GPUs/CPU clusters using Horovod package- Implemented models in PyTorch for benchmarking, served the model using OpenMP/MPI frameworks- Developed a novel deep neural network (DNN) architecture for thermodynamic phase identification (multi-class classification) and phase-split determination (regression) -Developed novel features that capture critical thermodynamic phase behavior resulting in high accuracy of the model2. Initiated and led a project to develop machine-learning based models for crude-oil property prediction- Performed extensive exploratory data analysis on data gathered from disparate sources, used clustering techniques to segment the data and employed ensemble methods as well as deep-learning for property prediction- Performed hyperparameter tuning, improved prediction accuracy and reduced inference time- Patent granted, "Machine-learning-based models for phase equilibria calculations in compositional reservoir simulations", App# US158797933. Initiated and led a project on developing deep learning based models for detecting liquid accumulation anomalies in gas-producing fields:- Developed deep-learning models using LSTMs for detecting anomalies in time-series data of gas producing fields; Potential response time reduction by 6 months- Developed models using Spark Streaming and Spark MLlib for real-time data ingestion, machine-learning and inference- Patent granted, "Real-time monitoring of hydrocarbon productions", App# US165170254. Contributed to the collaborative development of an in-house, scalable, high-performance production level code in Python/C++ for performing large-scale reservoir simulations

Performed optimization of globally weighted equity portfolio using fundamental index theory and cluster analysisDeveloped an in-house index as per the specifications of sponsors having a better performance than S&P 500 and lower re-balancing costs; Presented a poster at MIT-Sloan “open-house” conference