I Have No Coding Skills. My Team Is 3 AI Models. Here's What We Built in 2 Months.

Bogdan — Tue, 06 Oct 2026 18:43:19 +0000

Two months ago I started building a SaaS with zero programming background. My whole team is 3 AI models and one human (me). This is the honest account — the features, the bugs only an AI could find, and the rule we now build by.

The setup

I'm a solo founder from Ukraine. No code background. Instead of learning to code first, I hired a team:

AI #1 writes features.
AI #2 handles design and infrastructure.
AI #3 exists for one purpose: auditing everything the first two ship.

I'm the founder, product manager, and QA department of one. We run three products (ThumbRank, GrantPro AI, DataEmpire) with two more in development.

The problem we solve

Every YouTube creator knows the math: a viewer scrolling search results spends about half a second on each thumbnail. If yours doesn't win that half-second, the video doesn't exist — no matter how good it is.

Yet almost everyone checks their thumbnail the wrong way: full screen, on a desktop monitor, alone. Not next to nine competitors. Not on a phone. Not in the actual place where the decision happens.

I spent three years growing a channel and made every mistake in that list. ThumbRank is the tool I wish existed.

What we built (stack: Next.js + Supabase + Vercel + Paddle)

SERP Radar — type any keyword, see the real top-9 YouTube results, every thumbnail scored for contrast, focus, and readability. Upload yours, get a verdict: clickable, or invisible.
AI Thumbnail Score — computer-vision grading in the browser (canvas + two-step downscale so JPEG noise doesn't move the score). Nothing is uploaded to a server.
Competitor tracking — follow up to 10 channels, get alerted when one swaps a thumbnail. A thumbnail swap is the most honest signal of what's working in a niche.
Channel X-ray — every week the system finds your three weakest thumbnails and writes a diagnosis for each.
Video Doctor — paste a link to any of your videos (even unlisted), get a full packaging report: title, thumbnail, hook, structure.
Morning digest — a daily email with new competitor videos and alerts. (Resend + verified domain, because emails that only reach the founder aren't a product.)

Month 1 looked like progress

180K+ previews rendered. Features shipping daily. A launch plan, affiliate links, the whole classic indie playbook.

Everything looked fine.

Everything was not fine.

Month 2: the audit

Before letting anyone pay, I pointed AI #3 at the codebase the other two had written. It read 124 files and came back with 14 bugs. Four were critical — and every one was invisible from the outside:

The payment webhook returned 401 for days. Root cause: the signature header format (t=...;h1=...) wasn't parsed correctly, and after rotating the secret every live event failed — while the dashboard politely showed "Active". My own subscription was broken and I didn't know it.
A cron job had never run once. Silently. Since the day it was created. The schedule existed, the code existed, the trigger didn't.
A wrong price ID sat under a correct-looking button. A real customer would have paid $49 for the $79 plan and received the wrong thing, with no error message anywhere.
Two ID systems. profiles.id ≠ auth.users.id — rows were linked to the wrong identity in half the app. Took a SQL migration (auth_user_id backfill via email) to converge.

The fix pattern that emerged: store webhook events first, dedupe by event ID, retry failures, and never trust a green dashboard. The dashboards were green the entire time.

The rule we now build by

Fix first. Sell second.

Then, because trust in a fix is not proof, I ran a real $20 charge through my own checkout. A fire drill, not revenue. The webhook fired, the plan flipped, the badge updated. Zero real customers so far — but the first one can't break what we already broke ourselves.

What's next

More honest numbers, posted publicly. The Product Hunt launch. The first real sale when it happens — and the reasons why it took as long as it did.

If you're building with AI as your team: the code it writes is only half the product. The other half is what you build to catch it.

Building ThumbRank in public. Free preview, no signup: thumbrankpro.com

DEV Community: Bogdan