ZYRA Guardian

Inspiration

Food labels are written for local shoppers, but allergies travel across borders.

When we travel, we often need to make food decisions from labels written in languages we do not understand. A simple translation is not enough: the user needs to know whether a selected allergen appears directly in the ingredients, only in a “may contain” warning, or whether the scan is too incomplete to trust.

During Build Week, I traveled to Warsaw and tested this problem using a real Polish product and a simulated profile for a traveler with a severe tree-nut allergy.

What it does

ZYRA Guardian is a personalized allergen decision layer inside ZYRA, an existing multilingual food-analysis platform.

The user selects allergens to watch for. After scanning a product, Guardian compares the structured ingredient result with the personal profile and produces one of four outcomes:

DANGER — a selected allergen was found directly in the ingredients
MAY CONTAIN — the label includes a trace warning
NO PROFILE MATCH — no selected allergen was detected, without claiming the product is safe
SCAN INCOMPLETE — the ingredient list was not captured reliably enough

Guardian preserves the original evidence from the package. In the Warsaw test, it connected the Polish phrase “orzechy laskowe 13%” to hazelnuts and the selected tree-nut profile.

The experience follows a simple principle:

Decision first. Evidence second. Analysis third.

What I built during Build Week

ZYRA existed before Build Week as a working multilingual food scanner. During Build Week, I created the new Guardian vertical slice:

a persistent personal allergy profile
personalized matching against existing structured allergen results
separate handling of direct ingredients and “may contain” warnings
fail-closed behavior for incomplete OCR
four clear Guardian decision states
visible source evidence from the original foreign-language label
mobile-first Guardian interface
Polish-language validation with a real product bought in Warsaw
regression tests protecting the existing scanner and allergen contracts

How I built it

ZYRA uses a Next.js and TypeScript frontend with a FastAPI and Python backend. Real package images pass through OCR, ingredient extraction, multilingual normalization and structured allergen detection.

Codex was used to audit the existing architecture, identify a safe integration point, implement the Guardian profile and deterministic decision layer, create tests, inspect regressions and harden the mobile interface.

GPT-5.6 is used meaningfully in ZYRA’s multilingual normalization workflow to help transform noisy ingredient data into clearer structured information.

The final allergen decision remains deterministic and traceable. GPT-5.6 does not independently invent, remove or downgrade an allergen warning.

Challenges

Real packaging is difficult: text may be small, curved, reflective, multilingual and mixed with marketing copy or nutrition tables.

The most important challenge was avoiding false confidence. Guardian must distinguish between:

an allergen directly present in the ingredients
a manufacturer’s trace warning
no detected match
insufficient evidence

For that reason, ZYRA Guardian never displays “SAFE” or “safe to eat.”

Accomplishments

Built and tested Guardian during a real trip to Warsaw
Detected a tree-nut match from a real Polish ingredient label
Preserved the exact Polish source evidence
Added personalization without rewriting the stable OCR pipeline
Kept direct ingredients separate from trace warnings
Created a working production-style feature rather than a staged translation demo

What I learned

AI is most useful in high-trust workflows when it improves understanding without hiding the evidence.

For food-allergy decisions, the system should show what was found, where it was found, how it maps to the user’s profile and what action to take next. It should never manufacture certainty.