ingredients.help
Food additives, in plain language
Look up any additive and read it in plain language — then drop two or more into the Laboratory to see how they actually interact. Some pairings are risky, some are helpful, and where nobody knows yet, the answer is an honest “unknown” — never a comforting guess.
The back of a food package is a wall of E-numbers and chemistry. Most “is this bad for me?” answers online are either fear-mongering or hand-waving. ingredients.help sits in the middle: a hand-curated catalog of food additives, each explained across five plain-language sections — Origin, Usage, Safety, Properties and Regulation — with real ADI values, IARC classifications and regulatory notes, written for a person, not a chemist.
But an ingredient rarely acts alone. The product’s real question is what happens when additives meet — and that’s the Laboratory.
The founding principle
No data never means “safe.” When a combination isn’t in the evidence, the app says unknown — out loud.
30 curated additives (E-numbers) across 8 categories — preservatives, sweeteners, colorants, thickeners, acidulants, antioxidants, emulsifiers, flavor enhancers — each with formula, concern level, aliases and the five-section explainer. Pages are statically generated.
Add two or more additives, press Analyze, and a deterministic engine checks 20 documented interaction rules — surfacing the single highest-risk finding, from harmful reactions to intentional beneficial synergies. Your bench persists locally between visits.
An exact-match resolver for names, E-numbers and synonyms sits on top of fuzzy search, so “sorbate” returns an honest ambiguous (E200 vs E202) rather than silently picking one. Typo-tolerant, never over-confident.
Sign in (optional) to save combinations and keep a recently-viewed history — each saved analysis frozen as an immutable snapshot, so it stays explainable even after the rules are updated.
Every result rides a four-step risk meter — Low → Moderate → Elevated → High — plus a distinct state for “unknown.” Two real examples the engine actually encodes:
Together, and especially with heat or light, benzoate and vitamin C can react to form benzene, a known carcinogen. Condition-gated, but flagged even when conditions are unstated — at reduced confidence.
Not every interaction is a warning. These two gums reinforce each other, giving more viscosity together than apart — an intended pairing, and the meter says so plainly.
A pure, unit-tested function evaluates 20 rules with pair, set and category scopes. Same inputs → same verdict, every time. No model in the loop, no black box — the reasoning is inspectable.
Rules can be gated on heat, pH, light or personal sensitivity. When context is missing, a gated rule still surfaces — at halved confidence — so a real risk is never silently hidden.
Every analysis stamps its engineVersion and
rulesVersion and freezes the ingredients it used. Old saved
results stay explainable even after the rulebook moves on.
The name resolver returns exact, ambiguous or
none — the seam a future label-scanner will lean on to avoid
confidently misreading a package.
Warm paper, ink-green text and a calm sage-and-mint accent — clinical enough to feel trustworthy, soft enough to feel human. Headings in Fraunces, body in Inter. The risk scale is a fixed, color-blind-considered ramp from sage to clay so a verdict reads the same way every time.
Next.js 14 (App Router · React 18) with TypeScript and Tailwind. The catalog and engine run fully client-side with no backend required; Supabase is optional, adding email/password accounts, saved combinations and view history. Zustand persists the Laboratory bench, Fuse.js powers fuzzy search, and Vitest covers the deterministic core. There is no AI in the product today — label OCR and generated explanations are deliberately marked as future work, not shipped as claims.
The catalog, search, Laboratory and the deterministic engine all work today, with the interaction core covered by unit tests. Accounts and saved history are wired through Supabase. What’s not built yet — and isn’t pretended to be: