跪拜 Guibai
← All articles
VibeCoding · AI Programming · ChatGPT

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

By 大侠Luffy ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A working, paid SaaS product built entirely by AI — including a custom infrastructure scheduler — resets expectations for what a solo developer can ship in under two weeks. The deliberate choice to expose MCP alongside a REST API signals a practical bet that AI agents, not just human users, are becoming a paying audience.

Summary

PennyScribe converts audio and video into clean, multilingual text using Qwen ASR, deliberately skipping timestamp generation to keep the pipeline simple and costs low. The service is designed for AI agents and developer workflows that need to feed transcripts into downstream summarization, extraction, or retrieval tasks. It exposes CLI, API, and MCP interfaces rather than locking functionality behind a UI.

Under the hood, the infrastructure runs on cheap consumer GPUs rented from Vast.ai. Because those instances are inherently unstable, the builder used Codex to scrap the official Serverless option and write a custom GPU instance scheduler from scratch. The entire product, including the scheduling layer, was produced without writing a single line of code manually.

What would have been a three-month solo build was compressed to under two weeks. The current bottleneck is GPU scheduling throughput, and further architecture work aims to drive down the per-hour transcription price.

Takeaways
PennyScribe is a paid audio transcription service built with zero manually written code.
Transcription runs on Qwen ASR, with GPU compute rented from Vast.ai's consumer-grade instances.
A custom GPU scheduling system replaced Vast.ai's Serverless option to balance cost and reliability.
Timestamps are deliberately omitted because Qwen ForcedAligner produced poor results, and fixing them added too much complexity.
The service exposes CLI, API, and MCP interfaces, targeting AI agents as first-class consumers.
Building the entire product took under two weeks; a manual build was estimated at three months.
Conclusions

Treating AI agents as a paying customer segment — not just human users — is a product thesis that justifies exposing MCP alongside a traditional API.

Omitting timestamps is a deliberate scope cut that avoids a known weak point in the model pipeline, trading a common feature for higher reliability and lower cost.

Building a custom scheduler on top of spotty consumer GPUs, rather than paying for managed infrastructure, is a cost play that only becomes practical when AI handles the integration work.

Concepts & terms
MCP (Model Context Protocol)
An open protocol that standardizes how AI agents discover and interact with external tools and data sources, analogous to a USB-C port for agent-tool communication.
Qwen ASR
Alibaba's automatic speech recognition model that converts speech to text, used here for multilingual transcription.
Vast.ai
A GPU marketplace that rents consumer-grade and datacenter GPUs at low hourly rates, but with less reliability than managed cloud providers.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗