Rahul Varma

A lightweight U-Net-based binary segmentation model that identifies salient regions in images, enabling efficient deep learning on resource-constrained edge devices by reducing unnecessary computation.

Technical Approach: QU-Net generates binary masks highlighting regions of interest, allowing downstream models to focus computation only on relevant areas instead of processing entire images equally.

Key Features:

• Binary segmentation for region proposals on edge devices
• 25% computational reduction on MS COCO dataset
• 57% computational reduction on Cityscapes dataset
• Outperforms Dynamic Convolutions baseline

Impact: Enables deployment of sophisticated computer vision models on edge devices with severe resource constraints, making AI more accessible for IoT, mobile, and embedded systems.

📄 Read Full Thesis →

Part of Collaborative Earth's initiative to support global reforestation efforts by making domain knowledge more accessible to practitioners worldwide.

Technical Approach: Developed a RAG-based search and retrieval system using vector embeddings to enable practitioners to find relevant reforestation information from curated knowledge sources.

Key Features:

• Scraped and processed data from open-source journals and publicly accessible resources
• Built vector database with embedding models for semantic search
• Implemented retrieval augmented generation (RAG) for Q&A functionality
• Developed automated data pipelines for continuous knowledge curation

Impact: Makes reforestation knowledge accessible to global practitioners, addressing barriers in finding actionable information for specific projects and supporting global ecological regeneration efforts.

An auto-scalable cloud application providing real-time object detection as a service using AWS infrastructure and the Darknet machine learning model.

Technical Approach: The system processes real-time video streams from Raspberry Pi devices, automatically scaling cloud resources based on demand to handle concurrent requests efficiently.

Key Features:

• Real-time object detection on video streams using Darknet ML model
• Automatic horizontal scaling (scale-out/scale-in) based on request volume
• Concurrent request handling with intelligent resource allocation
• Integration with AWS EC2, S3, and SQS for distributed processing

Impact: Demonstrates scalable cloud architecture for ML inference, enabling cost-effective deployment of object detection services that automatically adapt to varying workloads and demand patterns.

A neural network pruning framework that reduces model size and computational requirements while maintaining accuracy through intelligent structured pruning techniques.

Technical Approach: Implements L1-norm based structured pruning on ResNet architectures, systematically removing less important filters and channels to compress models for efficient deployment.

Key Features:

• Structured pruning using L1Strategy on convolutional layers
• Dependency-aware pruning that maintains model integrity
• Filter ranking based on activation outputs and gradients
• Iterative pruning with fine-tuning for accuracy preservation

Impact: Reduces model size and inference time while preserving accuracy, enabling deployment of deep learning models on resource-constrained devices and lowering computational costs in production environments.