Inspiration
Food packaging contains plenty of information, but understanding it is another task. Unfamiliar ingredients, nutritional values, and claims such as “zero sugar” can leave consumers with questions that a simple ingredient dictionary cannot answer. PackTruth aims to make this information understandable without fearmongering or false reassurance.
What it does
PackTruth is a local web application for interpreting food packaging. Users upload a package photo or screenshot, review and correct the extracted information, and explore an integrated analysis of ingredients, nutrition, and marketing claims.
The ingredient view explains normal food-use assessments, findings from higher-intake studies, evidence-based considerations for particular populations, and the relevance to the current food. An expandable view provides sources, study conditions, doses, and limitations. Composite ingredients can be expanded into their named components, while unspecified categories are not treated as known chemical mixtures.
Four existing fictional samples let visitors explore the workflow without an API key. Live image recognition and health research require internet access and the user's own DeepSeek API key. The key is kept in page memory; requests retain budget limits, cancellation, timeouts, and credential isolation.
How it was built
The frontend uses React and Vite, with an Express backend running on Node.js. DeepSeek supports image extraction and structured research tasks; the application orchestrates Search, Fetch, and Chat operations to retrieve and explain scientific evidence. Deterministic calculations and food-label rules complement the AI-generated explanations. WorkBuddy supported the main development process, with Codex assisting later reviews, fixes, testing, and packaging.
Challenges and lessons
A central challenge was moving beyond ingredient identity and technical uses to useful health explanations. The research flow must track aliases and mixture components, read relevant sections of long assessments, and preserve credible findings rather than discard an entire answer because part of the evidence is incomplete.
Equally important is keeping research conditions intact: a high-dose experiment is not a prediction about eating one particular product, and an acceptable daily intake is not itself a description of a bodily effect. The application does not infer additive quantities that are absent from the label.
What we are proud of
The project connects recognition, human confirmation, nutrition calculations, label interpretation, and traceable evidence in one workflow. Automated checks cover domain calculations, scientific safeguards, health-answer organization, and BYOK isolation. Offline demonstrations and mocked tests remain distinct from live research.
Running the project
The complete source code is publicly available under the MIT license on GitHub. The README explains local installation and startup. This is currently a computer-local application, not a publicly hosted service or a standalone mobile app. It provides food information and science communication, not diagnosis or treatment advice.
What's next
Possible future improvements include broader evidence coverage, clearer source navigation, and additional usability testing, while preserving transparent research conditions and user-controlled API spending.
Built With
- deepseek
- express.js
- javascript
- node.js
- react
- vite
