VISION AI

Nutrition Assistant

A label scanner that answers one question — can I eat this — for people managing a specific dietary constraint.

ROLEProduct Manager

PERIOD2019 — 2020

CONTEXTDatis E-health

THE PROBLEM

Nutrition apps ask users to log food. People with a real dietary constraint do not want a diary; they want a yes or no about the packet in their hand, in the time it takes to read a shelf label.

3 tapsthe budget for a decision made standing in a supermarket aisle

CONTEXT

The existing app had a logging flow with a 12% weekly retention. Interviews showed the actual job was pre-purchase, not post-meal.

APPROACH

Point the camera at the label, get a verdict against the user's stored constraint. Logging became an optional by-product rather than the price of entry.

MY ROLE

I killed the logging-first onboarding, defined the single-verdict interaction, and set the accuracy floor below which we show the raw label instead of a verdict.

HOW I WORKED

01InterviewsTwenty-two users with diagnosed constraints; the aisle moment came up in nineteen.

02Constraint modelSix constraint types encoded as ingredient rules rather than free-text preferences.

03Verdict designOne of three states — safe, unsafe, unreadable. No scores, no percentages.

04Field testTwo weeks of real supermarket use with instrumentation on failed scans.

DECISION LOG

The calls I owned, the alternative I rejected, and the cost I accepted for each.

DECISION

Show 'unreadable' rather than a low-confidence verdict

ALTERNATIVE REJECTEDAlways return a verdict with a confidence indicator

WHYA wrong 'safe' has a medical cost. Unreadable is an honest state that the user can act on by reading the label themselves.

COST ACCEPTED14% of scans return nothing useful.

DECISION

Drop food logging from onboarding

ALTERNATIVE REJECTEDKeep logging, add scanning as a feature

WHYLogging was the reason users churned in week one. The scan is the job; the diary was our idea, not theirs.

COST ACCEPTEDLost the longitudinal data set the analytics roadmap assumed.

DECISION

Three states, no numeric score

ALTERNATIVE REJECTEDA 0–100 compatibility score

WHYA score invites interpretation the model cannot support. Three states are checkable and unambiguous at a glance.

COST ACCEPTEDLess differentiated in a store listing full of scored competitors.

WHAT MOVED

BEFOREAFTER

Week-one retention12%41%

Actions to a verdict93

Sessions containing a log entry100%22%

OUTCOME

50K+users on the platform

+45%engagement after the scan-first rebuild

3taps from camera to verdict

WHAT I PRODUCED

Constraint rule modelVerdict state specAccuracy floor policyField-test instrumentation plan

WHAT I'D DO DIFFERENTLY

We validated the aisle moment thoroughly and the repeat-purchase moment not at all. Users scan the same products weekly; a remembered-verdict cache should have been in the first release.

NEXT PROJECTNLP Comment AnalysisComprehensive competitive analysis system analyzing reviews using NLP and sentiment analysis with advanced AI models.