Srecharan Selvam

I’m a recent MS graduate from Carnegie Mellon University specializing in machine learning, computer vision, and AI systems. My passion lies in building intelligent systems that bridge cutting-edge research with real-world applications.

At Carnegie Mellon, I work as a Graduate Research Assistant at Kantor Lab, where I developed a Vision-Language-Action (VLA) system through LoRA fine-tuning of LLaVA-1.5-7B foundation model, adapting multimodal reasoning from text generation to action prediction for robotic leaf manipulation. This research combines self-supervised learning, computer vision, and geometric algorithms, involving custom CUDA kernel development and deployment on a 6-DOF gantry robot. The work is currently under review at ICCV 2025.

My industry experience spans ML engineering roles where I built real-time 3D hand gesture recognition for AR-based training systems and developed distributed computer vision pipelines for industrial automation.

Through personal projects, I’ve explored diverse AI applications: developing Vision-Language Models with RAG and RLHF techniques, building algorithmic trading systems with ensemble ML and sentiment analysis, and creating synthetic data augmentation systems using GANs, VAEs, and diffusion models for data-limited scenarios.

I focus on end-to-end ML pipeline development - from self-supervised data creation through distributed training, optimization with CUDA/TensorRT, and production deployment using Docker and cloud platforms. My approach emphasizes building practical, scalable solutions that actually work in production environments.

selected publications

Self-Supervised Learning for Robotic Leaf Manipulation: A Hybrid Geometric-Neural Approach

Srecharan Selvam, Abhisesh Silwal, and George Kantor

arXiv:2505.03702, 2025

Under review at ICCV 2025

Abs arXiv PDF

Automating leaf manipulation in agricultural settings faces significant challenges, including the variability of plant morphologies and deformable leaves. We propose a novel hybrid geometric-neural approach for autonomous leaf grasping that combines classical computer vision with neural networks through self-supervised learning. Our method integrates YOLOv8 for instance segmentation and RAFT-Stereo for 3D depth estimation to build rich leaf representations, which feed into both a geometric feature scoring pipeline and a neural refinement module (GraspPointCNN). The key innovation is our confidence-weighted fusion mechanism that dynamically balances the contribution of each approach based on prediction certainty. Our self-supervised framework uses the geometric pipeline as an expert teacher to automatically generate training data. Experiments demonstrate that our approach achieves an 88.0% success rate in controlled environments and 84.7% in real greenhouse conditions, significantly outperforming both purely geometric (75.3%) and neural (60.2%) methods. This work establishes a new paradigm for agricultural robotics where domain expertise is seamlessly integrated with machine learning capabilities, providing a foundation for fully automated crop monitoring systems.