SuppScore.com: supplement analysis platform

The problem

Supplement labels are hard to compare. Two bottles of magnesium can contain different forms at different doses, with fillers and additives in the small print, and the shelf price says little about how much active ingredient you actually get. Dr. Vegan wanted shoppers to be able to check any supplement, not only its own, against the standards it uses when formulating its products.

What we built

SuppScore.com, a public analysis platform. A shopper can start from any of four inputs: a product URL, a barcode, a photo of the label, or ingredient text pasted straight in. All four end in the same place: a structured breakdown of what is in the product, a ranking against Dr. Vegan's ingredient standards, and a value score.

Alongside it we built an A–Z ingredient dictionary, one page per ingredient, written for search and marked up with structured data so search engines and AI assistants can read it.

How it works

• Getting the label. Product URLs go through a scraping pipeline that fetches the product page and finds the ingredient panel. Barcodes are matched to a product. Photos go through a vision model that reads the label text.
• Reading it. A language model turns the raw label text into structured JSON: each ingredient, its form, the dose per serving, and anything that is an additive. The output is validated against a schema before anything else uses it.
• Ranking it. The ranking is not done by the model. It is a deterministic engine in ordinary code that checks each ingredient against Dr. Vegan's standards, so the same product always gets the same result and every score traces back to a rule.
• Valuing it. The value score works out price per unit of active ingredient, not price per bottle, so a cheap product with a low dose does not look better than it is.

Outcome

Shoppers get a consistent, explainable answer for any supplement, whichever way they start. Dr. Vegan gets its own standards applied in public, and a dictionary that builds search traffic over time. Because the model only reads and the rules decide, the team can change a standard and re-rank every product without retraining anything.

Stack

Part	Tools
Web app	Next.js, React
API	Node.js, Fastify, TypeScript
Label reading and extraction	Google Gemini (vision and text), Zod schemas
Data	PostgreSQL, Drizzle ORM
Hosting	Google Cloud, Terraform

Next case studyA sales portal on Dynamics 365 for a packaging manufacturer