How GPU-Powered Coding Agents Can Assist in Development of GPU-Accelerated Software
This blog post chronicles how VS Code equipped with GitHub Copilot powered by Claude Opus 4.6 was used to port the open-source whisper-asr-webservice project to NVIDIA Jetson hardware with full GPU acceleration — navigating over 15 build iterations, compiling CTranslate2 from source for aarch64 CUDA, working around Poetry resolver conflicts and pip wheel priority bugs, creating runtime compatibility shims for torchaudio, torch.load, and huggingface_hub API changes, testing all three ASR engines with self-generated speech audio, and ultimately forking the repo, opening a detailed pull request, and pushing a pre-built container image to Docker Hub — all driven by natural-language prompts — demonstrating how GPU-powered AI coding agents can come full circle by building GPU-accelerated software for edge devices like the Jetson Orin, unlocking practical automations such as automatic subtitle generation for Plex media libraries via Bazarr integration.
