How we built infrastructure
Genera was one of the first architecture and home interior AI apps in 2023. It started as a small side project, hacked over a couple of weeks, but surprisingly found its audience quickly after the Product Hunt launch. Over the next few weeks, thousands of people signed up and used it, which felt pretty surreal at the time.
Before launching Genera, there were two stacks we needed to figure out: compute and financial. At the time, Vercel made it super easy to spin up lightweight, front-end–heavy apps. But we needed to deploy fine-tuned Stable Diffusion models, and that meant spinning up your own GPU machine on demand in the cloud. Fast forward only a couple of years, and now there are plenty of serverless options. At first we used lambda.ai, with a Node front-end that kept track of long-running jobs.
For financial infra, we just went with Stripe, but had to roll out our own billing system to accommodate credit-based billing. A $20/month plan gave users 1000 credits, which was our internal virtual currency, and then two processes consumed it: LLM inference behind the scenes to process users' inputs, and the image generation process. Now AI-native billing systems like Grant, Paid.ai, and Lago are available — they would've saved days of work back then.
If I were to redo this project in 2026, here's what I would do:
Compute stack:
- Vercel — for front-end hosting.
- OpenRouter for an inference gateway. That way we can easily experiment with all models and providers.
- Gemini / OpenAI / Chinese open-source image models. These days they are so powerful, you can extract so much through better prompting — no need for fine-tuning anymore.