Context
This was a speculative concept project exploring how AI could personalize food recommendations using biological data rather than stated preferences alone. Julia and I collaborated on every stage of this project together, from concept through prototype.
The Concept
TasteScan pairs a portable saliva-reading device with a companion app that builds a personalized "taste fingerprint," tracking sensitivity to salt, bitterness, sweetness, and spice, then uses that fingerprint to recommend food and drinks via menu or barcode scanning, with a "Flavor Fit Score" shown through a simple green/yellow/red system.
How it would work: the saliva reader uses biochemical and electrochemical sensors to detect specific compounds tied to taste sensitivity: sodium levels (salt), glucose (sweetness), bitter-compound analogs like quinine, and capsaicin perception (spice tolerance). We saw this as valuable beyond just better food recommendations. The same biological data could help someone recognize patterns relevant to their health, like consistently high sensitivity to a compound that correlates with a dietary intolerance, without needing a clinical test to notice it.
Why We Designed It This Way
We mapped out core user groups (chefs, health-conscious eaters, adventurous diners, sommeliers) and their recurring pain points: palate fatigue, food regret, wasted money, and decision paralysis when facing an unfamiliar menu. The insight we designed around: stated food preferences are unreliable, but biological taste sensitivity is consistent, so a system built on the latter could, in theory, make meaningfully better recommendations than existing food apps.
We also designed MoodMatch, a secondary personalization layer letting the system factor in a user's emotional state (stressed, adventurous, tired) alongside their biological taste profile, so the product stays useful even on days without a fresh saliva scan.
Prototyping
We moved beyond a purely digital concept by 3D-printing a physical model of the saliva-reading device, to test the scale and form factor of what we were designing for, not just how the app screens would look, but whether the physical object made sense in someone's hand.
Design Decisions
- A color-coded Flavor Fit Score (green/yellow/red) rather than a raw percentage, to make the recommendation legible at a glance while scanning a real menu
- Optional mood input rather than mandatory, so the product still functions on days a user doesn't want to engage deeply
- A reinforcement-learning feedback loop ("Loved it" / "Too salty" / "Too bland") designed to let the system improve without requiring a new saliva scan every time
Reflection
Because this stayed at the concept stage within a single course cycle, the most important next step we didn't get to was real user testing, specifically, validating whether the Flavor Fit Score actually correlates with what people report enjoying. Without that, the core assumption behind the whole product (biological data predicts satisfaction better than stated preference) remains untested. If I were to continue this project, that would be the very first thing I'd build toward.