Zero Lines of Code: A GPU Scheduling System and a Paid Transcription API Built by AI

This is my first attempt to rely entirely on AI programming tools to develop a product from scratch. Throughout the entire process, I didn't write a single line of code myself.

The transcription model uses Qwen ASR, and the GPU computing power comes from Vast.ai. Although Vast.ai's consumer-grade GPUs are cheap, turning them into a stable and reliable online service is not an easy task.

To balance cost and service stability, I used Codex to abandon the official Serverless solution and built a GPU instance scheduling system from scratch.

I stepped into many pitfalls during this process, but it was very worthwhile. Without AI's help, I estimate it would have taken at least 3 months to do this well, but now it took less than two weeks—of course, for many experts, this speed might still not be considered fast.

Next, let me formally introduce my new product: PennyScribe.

PennyScribe is an audio transcription service designed for AI Agents and developers. It can quickly convert speech from audio and video into structured text, which can then be handed over to an Agent for content summarization, knowledge extraction, information retrieval, and automated processing.

If you've ever asked ChatGPT or Claude Code to handle audio transcription tasks, you might have found that they often need to download and run Whisper first. This not only makes the processing time very long but can also be interrupted due to insufficient device performance or network issues.

This is where an Agent-friendly cloud speech transcription service comes in handy. Especially when you need to process a large batch of audio files, the advantage of parallel processing with multiple GPUs in the cloud becomes even more apparent.

During testing, I found that Qwen ASR's text transcription results are excellent, but the timestamps generated by Qwen ForcedAligner are not ideal. Removing background music and environmental noise improves things, but it also increases the complexity and computational cost of the processing pipeline.

Therefore, PennyScribe has decided to focus on one thing for now: transcribing speech into high-quality plain text.

Although it's temporarily not suitable for video subtitling, if you need to convert interviews, podcasts, calls, or lecture recordings into clean multilingual text and then hand it over to an application or AI workflow for further processing, PennyScribe will be a very suitable partner.

PennyScribe offers CLI, API, and MCP, making it natively friendly for AI Agents.

This is also a bit of my thinking on the form of future new products: products should not only be convenient for people to use but should also allow AI to easily discover, understand, and invoke them, rather than locking all capabilities inside a UI.

Currently, there is still a lot of room for optimization in GPU scheduling efficiency and transcription throughput. Next, I will continue to optimize the architecture, reduce costs, and further compress the final price.

Of course, I'm not sure if there is still enough commercial space for this—after all, there are already many mature speech transcription products on the market. But I still want to build it and see how far a transcription service truly designed for AI Agents can go.