I tackled the challenge of deploying a heavy Vision Transformer (RT-DETR) on a resource-constrained NVIDIA Jetson device, achieving a 3.6x inference speedup (9 → 33 FPS). This optimization was a key factor in securing a ₩12B government contract and proved the commercial viability of using advanced ViT models in edge environments.

#PerformanceOptimization #EdgeAI #TensorRT #PyTorch #NVIDIAJetson

To resolve a critical development bottleneck, I architected 'PIPER', a standardized, asynchronous system for large-scale AI development. This system dramatically reduced data processing time from 4+ hours to under 5 minutes and streamlined new model integration, boosting overall team productivity.

#SystemDesign #MLOps #Automation #Redis #Python

I actively contribute to the open-source community, focusing on improving the stability and usability of state-of-the-art AI models for developers worldwide.

• RT-DETR: Enhanced TensorRT inference stability and added support for next-gen hardware. (#480, #620)
• Co-DETR: Improved developer experience by fixing Docker environment issues. (#123)
• EasyOCR: Corrected a critical bug in text detection coordinates. (#920)

#Community #Collaboration #ProblemSolving