💩 Crap-O-Meter - an AI to judge AI-generated products (because apparently that's where we are now)

The Problem

With 100K+ absurd products in the database, I was drowning. Manually reviewing products meant I can publish maybe 1K/month. At that rate, I'd finish sometime in 2037.

Worse, users kept generating brilliant stuff like "Business Meeting Shortener" alongside complete garbage like "asdasd" or "Bitter Smart" (just... random words with zero intention). I needed a way to separate the genuinely creative absurdity from the actual nonsense.

The Solution: Crap-O-Meter

Built a full AI evaluation pipeline called the Crap-O-Meter that scores every product on three dimensions:

• Text Coherence (0-10): Does the description actually develop the title concept?

• Image Relevance (0-10): Does the image show what the title promised?

• Creativity Factor (0-10): Is this hilariously absurd or just boring?

The system uses multiple Cloudflare AI models in a smart pipeline:

• LLM judges for semantic understanding

• Embedding models for similarity scoring

• Vision models for image analysis

• Image captioning with fallback strategies

Technical Highlights (for the nerds)

Bedrock & Pillars architecture: Foundation analysis runs first (gibberish detection, concept extraction), then context-aware evaluations build on that. If the title is just keyboard mashing, kill switch activates immediately - saves ~$0.0015 per garbage product.

Smart caching: Title and description captures are reused across evaluations. No duplicate AI calls. Saves 40-50% on API costs.

Two-pipeline image system: Caption-first approach (fast, cheap) with vision model fallback (slow, expensive). Tries 4 different API payload formats because Cloudflare's APIs are... creative.

Synonym-aware scoring: Knows that "tube" could mean toothpaste container, "bottle" is beer, "alien" is monster. Makes scoring actually work.

The Results

Now every product gets a quality score displayed as an animated gauge (0-100 scale). The system is surprisingly good at:

• Catching refusals: "I cannot generate content..." automatic score of 1.0

• Rewarding specificity: "Gasoline Powered Turtleneck" scores high, "Faceless Horror" scores low (too vague)

• Detecting true creativity: "Cat-To-Human Translator" = 9/10 creativity, "Smart Toaster" = 3/10

Products like "Shit On A Stick For Chefs" now correctly score high (7-8 creativity) because.. it's a good product? Meanwhile, abstract nonsense like random word combinations get brutally low scores.

The Meta-Irony

I'm using AI to judge the quality of AI-generated products on an AI-powered parody store.

It's AIs all the way down.

And honestly It works better than I expected. The system distinguishes between "creative absurdity" and "actual garbage" with surprising accuracy.

What's Next

Using these scores to:

• Auto-publish high-quality products (score > 7.5)

• Surface the best products for the homepage

• Filter the catalog intelligently

• Eventually: let users filter by quality scores

Still spending way too much on Cloudflare AI bills, but at least now I can automatically curate the chaos instead of drowning in it.

The most pointless project continues to require increasingly sophisticated engineering. Not sure if I'm optimizing for success or just polishing a very elaborate joke.