666 supplement reviews later — what I learned getting LLMs to actually fact-check supplements' marketing claims

Paste a supplement product URL into HypeCheck, get a teardown that traces
every claim back to a specific study, the dose used in that study, and
the dose in the product. Built it after almost buying too many "miracle"
supplements I'd seen advertised online.

Live at https://hypecheck.io. ~666 products and ~460 ingredients in so far.

The interesting problem wasn't the AI

Modern LLMs are already pretty good at refusing obvious nonsense if you
just ask them. The actual hard problem was dose math.

Example: a fiber product claims "supports weight loss." Psyllium husk
genuinely has been studied for satiety effects — at 10-15g per serving.
The product contains 2g. The claim technically traces to real research,
but at the dose offered, nothing happens.

That's where most marketing BS lives. Not in outright lies — in the gap
between "ingredient X does Y" and "this product, at this dose, delivers Y."

So HypeCheck pulls actual study doses from PubMed, compares them to the
doses on the label, and shows you the chain:

"Claim traces to [study], which used [dose]. Product contains [different
dose]. Here's whether those match."

The point isn't to tell people what to buy. It's to make the reasoning
behind every claim visible enough that you can judge it yourself.

Where it is today

• Product reviews: ~666
• Researched ingredients: ~460
• Pages in sitemap: ~2,000
• Weekly visitors: double digits 😅
• Revenue: $0

Stack: Python + FastAPI, Claude Sonnet + GPT-4, real PubMed retrieval
(with a post-hoc PMID verification pass — caught hallucinations in 3-5%
of early outputs).

Two things I'd love feedback on

1. Does the claim chain actually land? Pick a supplement you have real
   opinions on. Run it through. Does the reasoning HypeCheck shows you
   read as convincing, or does it still feel like "AI said so"?

2. Is this a problem real people care about, or just one I personally
   cared about? If you're not supplement-curious, what would have made
   you click this post?

Link: https://hypecheck.io

Brutal feedback welcome — most useful kind at this stage.