Working Experience

• Optimized Vision-Language Models (VLMs): Fine-tuned VLMs for visual question answering and reasoning tasks, achieving a 30% accuracy improvement in a binary QA task through LoRA fine-tuning on A100 40 GB GPUs.
• Explored Deep Contrastive Learning: Designed and implemented approaches for image similarity search using metric learning techniques (Siamese networks, proxy-anchor loss, and triplet loss) to improve accuracy and efficiency.
• Hand Grasp Classification: Built a PyTorch and Mediapipe-based model to classify seven distinct hand grasps in packing stations with ~90% accuracy, leveraging advanced video analysis for operational efficiency.
• Optimized YOLOv7 for Hailo Hardware: Improved YOLOv7 object detection inference pipeline through quantization, achieving real-time performance (30 FPS) while maintaining accuracy and confidence levels.
• Developed Efficient ROS2 Inference Node: Created a ROS2 node using ONNX runtime, optimized for NVIDIA GPUs with CUDA and TensorRT, enabling efficient model inference in resource-constrained environments.

• During the summer of 2022, I worked under the mentorship of Weaviate engineers:
- Developed the text summarization module for Weaviate, which helps users to summarize search results.
- Created the inferencing engine using FastAPI and the Hugging Face Transformers library.

• I contributed to the Sysco Warehouse Management System (SWMS) under the SWMS NewUI team, focusing on modernizing the classic SWMS application:
- Fixed UI-related bugs using React JS.
- Improved the service layer performance by re-implementing logic and optimizing queries using Java.

• Developed a 3D image conversion system leveraging AWS lambda container image support.

Teaching Experience

CS3042: Database Systems, University of Moratuwa

CS3953: Technical Writing, University of Moratuwa

Tech Talks

LiveRoom Tech Talks

CSE, University of Moratuwa

CS & ES, University of Moratuwa