Audry Hsu is a software engineer at RunPod and a co-creator of Fána, an open-source feature-management platform. She builds tools that simplify difficult infrastructure decisions, from controlling software releases to running AI applications on cloud GPUs.
Before moving into full-stack engineering, Hsu worked at Deloitte as a senior consultant focused on regulatory and data management. She subsequently helped develop Fána with three other engineers, creating audience-targeted feature flags that enable selective releases and rollback without reversing entire deployments. Its architecture supports reusable audience rules, real-time updates, self-hosting and managed deployment; pre-evaluated responses and separate browser and server credentials help prevent sensitive targeting data from reaching client applications.
At RunPod, Hsu focuses on making GPU-backed AI development more accessible. Her work emphasizes practical choices developers must make about hardware availability, scaling, responsiveness and cost.
- Remote GPU execution without leaving the IDE. RunPod Flash lets developers decorate asynchronous Python functions for execution on cloud GPUs while keeping application logic and debugging local. Hsu demonstrated switching between image-generation models optimized for speed and visual quality without repeatedly rebuilding containers or reprovisioning hardware.
- Serverless inference with explicit tradeoffs. In a GPU-backed endpoint demonstration, Hsu configured H100 GPUs with A100 fallback, worker limits and always-active capacity. She distinguishes persistent environments for experimentation, autoscaling workers for variable demand and clusters for larger training workloads, emphasizing the relationship between cold starts, availability and cost.
- Multi-model application orchestration. Hsu connected Qwen 3 for prompt generation, DreamShaper for image creation and Google’s Nano Banana 2 for composition. The pipeline illustrates how local Python can coordinate specialized models while moving computationally intensive steps onto remote infrastructure.
