SACASE
02FitFuel AI
2025—26CLOSE CASE ×
02 / 07Health / Applied computer vision2025—26

FitFuel AI

A nutrition product that treats computer vision as an estimate—not an excuse to hide uncertainty.

ROLE
Product design · Full-stack engineering · AI integration
STATUS
Product + AI pipeline
STACK
Next.js · FastAPI · PostgreSQL · PyTorch · ResNet50 · Gemini
PRODUCT INTERFACE / 02Recognition result
Food analysis result with detected meal, nutrition values, and portion controls
PRODUCT JOURNEY / 02

THE PRODUCT,
IN USE.

One complete interface state at a time. Each screen documents the task, the system response, and what the user can do next.

02 / ESTIMATE

Return a useful guess

Recognition and portion research produce a candidate food and estimated nutrition values.

01Recognition result
Food analysis result with detected meal, nutrition values, and portion controls
01 / CAPTURE

Start with the meal

A photo or manual search begins a food log in the user’s current day and goal context.

02Vision logging
FitFuel food logging page with vision scan entry point and recent foods
04 / GUIDE

Connect the day

The confirmed entry updates macros, adherence, meal context, and progress views.

03Nutrition progress
Nutrition adherence analytics with calories, macros, and recent trends
01 / CONTEXT

WHY THIS PRODUCT
NEEDED TO EXIST.

CHALLENGE

Nutrition apps can count calories, but a dependable beginner experience must connect goals, meal planning, food logging, progress, and uncertain image recognition without letting one model control the entire product.

RESPONSE

I separated deterministic nutrition logic from probabilistic model output. Food recognition and portion estimation can accelerate logging, while correction paths, user context, and rules keep the experience usable when confidence is low.

02 / OPERATING LOGIC

THE OPERATING
SEQUENCE.

  1. 01CAPTURE

    Start with the meal

    A photo or manual search begins a food log in the user’s current day and goal context.

  2. 02ESTIMATE

    Return a useful guess

    Recognition and portion research produce a candidate food and estimated nutrition values.

  3. 03CORRECT

    Keep the human in control

    The user can change the result, portion, or ingredients before the record becomes trusted.

  4. 04GUIDE

    Connect the day

    The confirmed entry updates macros, adherence, meal context, and progress views.

03 / PRODUCT DECISIONS

DECISIONS THAT
SHAPED THE BUILD.

01 / MODEL

Vision estimates, users correct

The UI exposes food and portion controls after recognition so uncertainty never becomes invisible product behavior.

02 / LOGIC

Rules own the plan

Meal planning and nutrition calculations remain deterministic; Gemini supports explanation and conversation rather than deciding the critical plan.

03 / PRODUCT

One daily operating view

Food, macros, water, steps, weight, and planned meals share one consistent tracking system for beginners.

04 / INTERFACE EVIDENCE

Secondary screens that extend the core journey without diluting it.

05 / 07Weight goal
Weight goal tracking page with trajectory and weigh-in history
06 / 07AI assistant
FitFuel AI assistant discussing meal substitutions and daily context
07 / 07Hydration
Water intake tracking interface with weekly trend
05 / MY RESPONSIBILITY

SCOPE OF
RESPONSIBILITY.

I designed the product experience and full-stack architecture, built the FastAPI/PostgreSQL foundation, and defined the PyTorch/ResNet50 direction for food recognition and portion estimation research.

The product proof covers dashboarding, vision-assisted logging, progress, weight, water, steps, and AI guidance. Model quality and portion estimation remain active engineering work rather than finished clinical claims.

  • Computer vision
  • Nutrition logic
  • Human correction
  • Beginner UX
OPEN SOURCE REPOSITORY
NEXT CASE / 03Project Recommender