How I Built a SaaS with 0 Coding Skills (Using Next.js + AI Vision)

#showdev #saas #ai #nextjs

I am not a "real" developer. I don’t have a CS degree. I don’t know how to center a div without asking an LLM.

I am a sports bettor who was tired of spending 5 hours a week typing data into Excel.

But last month, I launched mybets.gg—a fully functional SaaS with auth, a database, and an AI-powered computer vision pipeline.

Here is how I went from "Idea" to "Shipped" using AI as my co-founder, why standard OCR failed me, and the tech stack I stumbled into.

The Problem: The "API" That Didn't Exist
I wanted to build a bet tracker. The problem is that sportsbooks (especially the niche European or Crypto ones I use) do not have public APIs. They don’t want you scraping their data.

If you are a real dev, you might try to reverse-engineer their internal endpoints or build a Puppeteer scraper. But I’m not a real dev. And also, those scrapers break every time the site updates its CSS.

I needed a "Universal Adapter." Something that didn't care about the code, but looked at the screen.

Attempt 1: Tesseract.js & Regex Hell ❌
My first attempt was using standard OCR (Optical Character Recognition) libraries like Tesseract.js. I thought, "Easy, just extract the text."

It was a disaster.

A betslip is unstructured data. Tesseract would spit out: "Arsenal 1.90 $50 WIN"

I spent weeks asking ChatGPT to write complex Regex patterns to parse this. But as soon as I tried a different bookmaker, the layout changed: "WIN - Arsenal @ 1.90 (Stake: $50)"

The Regex broke. I was playing whack-a-mole. I almost quit.

Attempt 2: The "Vision" Pivot ✅
Then I realized: I don't need OCR. I need Understanding.

I scrapped the Regex. I switched to using Multimodal LLMs (like Gemini Flash and GPT-4o). Instead of asking the code to "find the text," I essentially sent the image to the AI with this prompt:

"Here is a screenshot of a sports bet. Extract the Team, Odds, and Stake, and return it as clean JSON."

It worked instantly.

It didn't matter if the background was dark or light.

It didn't matter if the "Stake" was at the top or bottom.

It didn't matter if the text was German or English.

The AI "understood" the context of a betslip. I had my Universal Adapter.

The "Citizen Dev" Stack 🛠️
Since I didn't have strong opinions on tech, I let AI recommend the stack. Here is what we built on:

The Framework: Next.js (App Router)
The UI: Shadcn/ui + Tailwind
The Database: Neon (Serverless Postgres)

What I Learned (The "Indie Hacker" Takeaway)

I knew exactly what the product needed to do because I was the user. I knew the pain of "Spreadsheet Hell."

If I had waited until I learned React properly, I never would have started. Instead, I treated AI as a "Senior Dev" who sat next to me. I explained the logic, and it wrote the syntax.

Did we write spaghetti code? Probably.

Is it optimized? Maybe not.

Does it work? Yes.

The Result