MSLS_V5_15 · Developing Software as a Product · FS26

Food Diary

AI-powered meal logging: clean code from day one

React Native FastAPI Supabase LogMeal AI Railway

Built with AI from commit one

My AI teammates

Perplexity

The Researcher

Research
API documentation
No code generation

Claude

The Architect

Planning
Architecture
Issue design

GitHub Copilot

The Builder

Implementation
Code generation
Reviewed on every PR

What it does

Goal

📷

Photo logging

Photograph food → AI identifies → calories logged automatically

✏️

Manual logging

Search ingredients → select amounts → save to diary

📅

Daily diary

Entries grouped by date with total daily calorie tracking

Why a new project?

A different starting point

TYPICAL APPROACH

Existing messy project
Retrofit clean practices
Before → after story

THIS PROJECT

Empty repository
Every practice from commit one
No before, only after

Day zero

Before writing any code

01 · Tech stack

React Native + Expo
FastAPI (async, typed)
Supabase (auth + db + storage)
Railway (deployment)

02 · Architecture

Proxy pattern
API key stays server-side
Mobile never calls AI directly
Clean separation of concerns

03 · Workflow

main = protected release branch
dev = integration branch
PR-only · Copilot review required
[log:slug] on every commit

Architecture decision

Chosen design: secure API proxy

Mobile app

React Native + Expo

Backend API

FastAPI routers + middleware

External services

LogMeal + Supabase

Factual design choices

Backend exposes /auth, /diary, /profile, /logmeal routers
Mobile never calls LogMeal directly
API keys stay server-side in backend environment
Supabase handles auth, Postgres data, and image storage

Why this architecture

Security: no third-party secrets in mobile app
Control: backend can enforce validation and error mapping
Flexibility: swap LogMeal integration without app-side changes
Operations: one CI path for lint, tests, and deployment

Git workflow in practice

Branch strategy with PR merges

main

Protected release branch. Merge target after validation.

dev

Integration branch. Work lands here through pull requests.

test / review

Short-lived working branches. PRs are reviewed and merged into dev, then main.

Enforced flow

Issue branch → PR → CI checks → review → merge

Project structure

From scaffold to full project

Day 0 3 folders · 3 files

FOOD-DIARY [GITHUB]
▶  backend
▶  docs
▶  mobile
    .gitignore
    AI_USAGE.md
    README.md

Now 7 folders · 13 files

FOOD-DIARY [GITHUB]
▶  .githooks
▶  .github
▶  .vscode
▶  backend
▶  demos
▶  docs
▶  mobile
.gitignore
AI_USAGE.md
CHANGELOG.md
codecov.yml
CONTRIBUTING.md
DEV_LOG.md
LICENSE
LIMITATIONS.md
package-lock.json
PRODUCT.md
PROJECT_STATE.md
README.md
setup.sh

▶ from Day 0 ▶ new folders ● scripts / changelog ● documentation ● config / lock

What was applied

Engineering practices

✅ Repository hygiene

README · LICENSE · Issues · branch protection

✅ Git workflow

main/dev branches · PR per issue · commit convention

✅ Clean code

Modular · type-safe · zero debug statements

✅ IDE setup

.vscode/ · ESLint · Prettier · Ruff · launch configs

✅ Testing

pytest · Jest · business logic covered

✅ Profiling

cProfile → flamegraph → decision documented

✅ CI / CD

Lint · test · coverage · deploy on every PR

✅ Virtual envs

Dockerfile · devcontainer · pyproject.toml · lockfile

Testing strategy

Coverage split by code area

Backend

40.0%

209 / 523 lines covered

Frontend

7.4%

59 / 798 lines covered

Whole code

20.3%

268 / 1321 lines covered

Coverage source (local reports)

Backend: backend/coverage.xml
Frontend: mobile/coverage/lcov.info
Combined value is weighted by total covered lines.

Performance profiling: POST /logmeal/segment

Call stack breakdown

100 %

              app.routers.logmeal.segment
            
3,547 ms
95 %

              httpx.AsyncClient.send [LogMeal API call]
            
3,375 ms
87 %

              ssl.SSLSocket.read ← HOTSPOT
            
3,091 ms
Network I/O wait · LogMeal ML inference time · not Python CPU
8 %

              StorageFileAPI.upload [Supabase]
            
291 ms
5 %

              FastAPI + Python application logic
            
165 ms

Key finding

The 87% hotspot is network I/O, not Python code. The LogMeal API's food-recognition model processes the image server-side. Python execution (including Supabase upload) represents only 5% of total latency.

Profiling result

Why only LogMeal was profiled

87%

LogMeal API

ssl.SSLSocket.read
~3,091 ms

Supabase Storage

non-blocking
~291 ms

Python code

FastAPI + logic
~165 ms

Why profile LogMeal?

87% of wall time blocked in ML inference on external servers. Decision: external bottleneck cannot be optimised from application code. Mitigations in place: 30s timeout, 429 proxying.

Why not profile others?

Postgres: not a bottleneck at current data scale. Storage: non-blocking, async. Python code: 5% of total, yields ≤165 ms max, not worth optimising.

CI/CD evidence from GitHub Actions

All required checks are green

Push checks: Backend, Mobile, AI usage check, deployment

Pull request checks: Backend, Mobile, AI usage gate

Backend job: Ruff lint, format check, pytest, CodeCov upload

Mobile job: ESLint, Prettier check, Jest with coverage, CodeCov upload

Policy job: Verify AI_USAGE.md entry exists for each commit

Continuous integration and deployment

What runs on every pull request

CI: passing Coverage: ~20% AI usage: enforced

Standard pipeline

→ Ruff lint + format check
→ ESLint + Prettier check
→ pytest with CodeCov upload
→ Jest with CodeCov upload
→ EAS Build → Android APK
→ Railway deploy on merge to main

AI enforcement

→ Pre-commit hook reads message
→ Extracts [log:slug] token
→ Checks AI_USAGE.md for entry
→ Rejects commit if missing
→ GitHub Actions repeats on PR
→ PR blocked without valid entry

Non-standard contribution

AI accountability: enforced technically

Every commit carries a log token

feat(#13) [log:copilot]: photo log screen
fix(#11) [log:manual]: bbox scale formula
docs(#33) [log:claude]: CONTRIBUTING.md

Hook rejects any commit without [log:slug]

Three tools, three purposes

Perplexity

Research · API documentation · no code generation

Claude (Anthropic)

Planning · architecture · issue design

GitHub Copilot

Implementation · code generation · reviewed on every PR

State at end

What was delivered

issues closed

100%

CI checks passing

~20%

unit test coverage

v1.0.0-beta

GitHub release + APK

✓ Working Android APK (EAS Build)

✓ README + Architecture doc

✓ Deployed FastAPI backend (Railway)

✓ User manual + Profiling report

✓ Full CI/CD pipeline

✓ CHANGELOG + CONTRIBUTING.md

✓ GitHub Release with APK attached

✓ All five user flows functional

Mobile app demo

Core user flows

Auth

Diary

Profile

Manual log

Photo log

AI segmentation

Questions

Architecture · Codebase · Profiling · AI usage · CI/CD

GitHub repo CodeCov Architecture doc Flamegraph SVG