Projects

Mental Health Meme Classification

A multimodal NLP course project to classify anxiety and depression symptoms from internet memes using Vision-Language Models.

2025

Addressed single-label (anxiety) and multi-label (depression) classification from memes as part of NLP coursework.
Augmented dataset by extracting OCR text and semantic triplets (e.g., Cause-Effect, Mental State) using the QWEN-2.5-VL-7B model.
Enhanced a reference architecture (M3H) with visual feature maps and fine-tuned a MentalBART model for classification.
Achieved competitive performance with a 65% Macro F1 score for anxiety and 63% for depression.
Developed an end-to-end inference pipeline with a Streamlit UI for interactive visualization of results.

PythonPyTorchHugging FaceStreamlitNLPMultimodal AI

Benchmarked and monitored a complex microservices application on Docker Swarm and GKE to evaluate performance and observability.

2025

Deployed the Death Star Social Network across local Docker Swarm and cloud-based Google Kubernetes Engine (GKE) environments.
Benchmarked three distinct configurations: single-node/single-replica, multi-node/single-replica, and single-node/multi-replica.
Integrated a two-tier observability stack using Pixie for real-time visualization and Prometheus for fine-grained metrics collection.
Analyzed performance trade-offs in latency, resource utilization, and scalability across different deployment strategies.

DockerKubernetesGKEPrometheusPixieMicroservicesObservability

Runner-up project at the KAVACH-23 National Cybersecurity Hackathon, building an end-to-end ANPR and Face Recognition system.

2023

Led a team of six and collaborated with Ahmedabad West traffic police for high-definition video data collection.
Engineered a decoupled API using YOLOv8 for detection, achieving 92% precision and 91% recall on number plates.
Developed a cross-platform React Native application for real-time monitoring on edge devices.
Explored multiple face recognition approaches, including deep learning (Siamese networks) and classical methods (dlib).
Finished as a national finalist (runner-up) in the KAVACH-2023 hackathon among the top 100 teams.

PythonPyTorchReact NativeYOLOv8Computer VisionANPR

State-level hackathon winner and published research on predicting student dropouts using machine learning and data visualization.

2023

Won the SSIP-22 State Level Hackathon by developing a dashboard and predictive analytics platform.
Created a data pipeline using official government data from UDISE+ to analyze dropout trends.
Implemented regression models (Linear, Polynomial) achieving a high R² value of 0.9976 on the custom EduDropX dataset.
Extended the project into a research paper published in the IEEE I2CT 2024 conference.

PythonPandasMachine LearningData VisualizationScikit-learn

A centralized portal for academic resources that attracted over 3,000 visits in its first week and now ranks top on Google.

2022

Identified the need for a unified platform and developed a centralized website to host scattered academic materials.
Organized resources into an intuitive structure by subject and semester, simplifying access for students.
Achieved significant user engagement with 1,000+ unique visitors within the first week of launch.
The platform has since become a go-to resource, enhancing the student learning experience.

Web DevelopmentSEOContent ManagementWeebly