ScanDrink

Inspiration

The idea for ScanDrink was born from recognizing the massive inefficiency and hygiene risks associated with liquor inventory in airline catering. Manual checks of partially used bottles rely on fallible visual estimation, especially with dark glass. This results in two major problems: high waste (discarding usable product) and serious compliance risks (violating strict Airline SLAs and hygiene rules). Our goal was to eliminate human error and guesswork by creating an objective and sustainable inspection system.

What it does

ScanDrink transforms the inventory process by using multimodal AI to deliver an immediate, actionable decision for every bottle.

Precise Content Analysis: It analyzes a photo of a bottle, prioritizing the reading of a digital scale to calculate the exact percentage of liquid remaining.
Instant Compliance: It applies a three-tier decision structure:Hygiene: Discard immediately if the product is Wine and the seal is broken.
Global Quality: Discard immediately if content is below 40%.
SLA Rules: For compliant bottles, it applies the selected airline's policy to determine the final action.
Promotes Sustainability: Promotes recycling actions when the final decision is REPLACE. It displays an awareness notice to encourage eco-friendly inventory management.

How we built it

The core of ScanDrink is a hybrid, multi-layered architecture:

A fast, mobile-friendly interface developed using React and Vite. It handles the camera capture, displays the real-time status (Processing, Analyzed), and shows the final recommendation.
AI Engine: We utilized the Google Gemini API (gemini-2.5-flash) as the central decision maker. We engineered a robust System Prompt that acts as a programmatic instruction set. It requires the AI to first perform visual analysis (identify product/category/seal), then locate and process the scale reading, and finally, calculate and output the percentage using its knowledge base (tare weight, gross weight).
Decision Logic: The main application logic parses the structured output from Gemini and translates it into the final action based on the hierarchical rules (Wine Rule > Global 40% Rule > Airline SLA).

Challenges we ran into

Inconsistent Visual Estimation: Initial testing showed wild variation in content estimates (e.g., 51% to 95% on the same photo) when relying only on vision, especially with dark bottles. Solution: This forced us to prioritize and implement the complex weight-based calculation prompt. We had to teach Gemini to be an "expert analyst" who researches the bottle's empty/full weight and uses the captured scale number as ground truth, significantly enhancing accuracy.

Accomplishments that we're proud of

Hybrid Accuracy Model: Successfully integrating vision and factual (weight-based) knowledge into a single and reliable output..
Prompt Engineering Mastery: Crafting a comprehensive System Prompt that functions as a programmatic script, handling complex conditional logic.
Compliance Automation: Translating regulatory documents (SLAs) into immediate, automated, and auditable actions, directly reducing risk for airline caterers.

What we learned

Trusting External Data: The primary lesson was that for tasks requiring precision (like volume or weight), AI APIs like Gemini are a powerful tool to teach how to optimize time in manual operations, by acting as an objective interpreter and calculator of external data (the scale reading) rather than a subjective estimator.

What's next for ScanDrink

Backend Enhancement & Authentication: Strengthen and improve our backend by increasing operational capabilities with the product catalog and implementing employee authentication for accountability.
Alternative Liquid Measurement: Implement other ways to identify the quantity of liquid a bottle holds, for example, developing a project using acoustics to determine liquid volume by analyzing the bottle's resonant frequencies.