The wild story of discode.ai's founder who spent $2k on APIs to generate a 417k-line mess with Claude, and their unique Eco-friendly AI router.
Imagine spending $2,000 on API tokens and writing a whopping 417,000 lines of code using Claude, only to realize you need to hire actual human software engineers to rewrite the entire thing from scratch.
No, this isn't a Silicon Valley sitcom. This is the actual origin story of discode.ai, an AI Router that recently made waves on Product Hunt.
It all started with Moriz, a coffee-house operator from Vienna, Austria. Moriz openly admits he can't code. His last encounter with programming was a basic HTML course back in the late 1990s. But powered by sheer curiosity and the AI hype train, he decided to build something anyway.
Enter "vibecoding"—the art of sipping espresso while asking an AI assistant to write software for you. After buying three Claude Max subscriptions and burning through $2,000 in API tokens, Moriz managed to build a massive, fragile, 417,000-line monster.
Realizing his creation was held together by digital duct tape, Moriz did what any pragmatic founder would do: he hired a team of professional engineers and designers to butcher that monster and rebuild it into a sleek, functional product called discode.ai.
So, what actually is discode.ai?
The Product Hunt crowd loved the concept but immediately started poking holes in the logic, especially regarding the eco-friendly metrics.
One sharp dev raised a very solid concern:
"Love the eco angle, but what if the cheap model under-delivers, the user gets frustrated, re-asks, and now you've burned MORE CO₂ than if you had just run GPT-4 once?"
Pete, one of the creators, owned up to this limitation. Since the tool is in beta, they aren't hiding retry costs—every single attempt is metered and displayed. They are currently working on a smarter "escalation" feature to automatically bump a weak response up a tier without requiring a manual, wasteful retry.
Another technical hurdle brought up was context memory:
"When multiple models are dispatched across a single conversation, how is memory handled? Does discode keep a central context layer, or does each model only see what is passed to it at that moment?"
Managing state across multiple API providers without losing context or blowing up latency is a tough nut to crack, and something the team is actively polishing.
There are two massive takeaways from the discode.ai launch:
First, vibecoding is great for prototyping, but it's terrible for production. Moriz's 417k-line nightmare proves that while AI tools can generate code at lightning speed, they lack architectural vision. You still need human developers to design clean, maintainable systems that won't collapse under their own weight.
Second, clever positioning is everything. While everyone else is fighting over who has the biggest parameter size, discode.ai focused on privacy and environmental impact—two massive talking points, especially in the highly regulated EU market.
Would you use an "eco-friendly" AI router, or is this just another greenwashing gimmick to get funding? Let us know in the comments below!
Source: Product Hunt
