Projects

Suraj has showcased his skills through diverse academic, professional, and personal projects, including open source contributions:

 Emotion Recognition in Multi-party Conversation using Acoustic Context

• Modeled an RNN-based architecture with context fusion for emotion recognition in multi-party conversation. Observed 3%

improvement in weighted-F1 score with models using acoustic and previous emotion contexts over the non-contextualized models.

Multi-modal and Multi-Hop Information Retrieval on WebQA dataset

• Crafted a multi-modal visual-linguistic fusion architecture that used the CLIP model to extract textual and visual representations

(from image and image captions) and projected them into a common representation space. Defined a custom batching strategy that

generates in-batch negatives to bolster learning.

• Evaluated against two-stage unimodal re-ranking models and observed 56% improvement in Recall@100 with our fusion model.

Spatio-Temporal Causal Discovery Framework for Hydrology System                                                                                                                                                      

• Worked on Spatio-Temporal Causal Discovery Framework, improving causal relationship identification by enforcing spatial and temporal constraints for water terrains. Collaborated with the product team of NSF to define the requirements.

• Predicted flow rates across different years outperforming traditional causal models.

SQL Query Optimization [Open-Source with IBM] | scikit-learn, SQL                                                                                                              

• Collaborated with innovative in-database machine learning data preprocessing to eliminate unnecessary data movement. Decreased preprocessing time by 70% for large datasets.

Secure Hospital Management System

• Built full stack web application using Django, MongoDB, AWS.

• Developed blockchain payment gateway and chatbot for navigation. Hosted on AWS.

Real-time Event Detection (Open Source)

• Analyzed Twitter data. Identified clusters using KMeans.

• Visualized events in real-time with D3.js and Flask.

SQL Query Optimization (Open Source with IBM)

• Developed in-database ML data preprocessing, reducing time by 70%.

• Pushed computation into PostgreSQL/DB2, enabling 10x larger datasets.

Automation Report Tool

• Mapped bugs and deployed on AWS using MERN stack.

• Integrated with CommScope's MLISA tool.

High Level Automation Framework

• Created Python scripts to interact with various devices.

• Developed predefined functions using Robo Framework.

Person Re-Identification Model

• Designed Pytorch model for CCTV-based person re-identification.

• Achieved 92.4% efficiency on CROWDHUMAN dataset.