Year
2026
3 months
Expected
of behavioral intention
but only
actual use behavior
Why can't the camera tell the difference?
Visual input isn't enough
To better the accuracy, this model needed more data acquisition than just images: digital purchase data.
I chose email receipts as the data acquisition method for two main reasons:
Barcode scanning and manual logging depend on human input, which is unreliable since most people simply don't do it.
Grocery delivery and order APIs offer cleaner data, but only cover online shoppers, when the majority still shop in-store.
What does receipt data improve?
Receipt data narrows the guesswork downstream, improving two key steps in the processing:
Feature-based classification
Instead of the model choosing from thousands of possible food classes, it only needs to pick between items that were bought recently.
Post-processing
If the classifier returns a low-confidence or ambiguous guess, receipt data can be cross-checked against what's expected to be in the fridge and nudged toward the more likely match.
Foodie: a license model for Samsung SmartThings, training Vision AI with digital purchase data.
The opportunity
There is a huge installed base of smart fridge hardware that is collecting dust in homes.
US Smart Refrigerator market
Scalability
Scalability and growth reinforcing each other rather than competing for resources.
I pitched Foodie in front of founders and investors in NYC!
Foodie placed 3rd and I had the chance to connect and get valuable feedback. :D
Always pivot
Foodie's first version was a B2C product bundled with its own camera hardware. Midway through building the prototype, I realized the business model wouldn't hold up. Customer acquisition costs alone would sink it. Pivoting multiple times taught me that this will always bring a greater outcome than continuing a product you already know will fail.
