to my petal — Erti Hoxha

Plant journal · latest projectto my petalAI-assisted plant care and visual journaling

Notes on giving a leaf photo an honest second look — and saying “I’m not sure” when that’s the truth.

p. 01 — the idea

Write your plants a love letter.

I built to my petal to help plant owners understand possible leaf issues and keep track of their plants over time. It combines image analysis, care reminders, and a photographic journal in a warm, botanical interface.

Photograph a leaf, get an honest second look, leave yourself notes, and watch each plant’s story grow.

p. 02 — the challenge

One leaf photo rarely tells the whole story.

Symptoms can look alike, growing conditions matter, and models trained on controlled photographs can struggle with real garden images.

The challenge: make image analysis useful while communicating its uncertainty clearly.

p. 03 — the solution

A journal that remembers how it looked.

A responsive, full-stack app where you manage your plants, upload leaf photos, add context about watering and light, and save findings to a personal journal.

Dated photo comparisons show change over time, while recurring reminders and source-backed care guides support ongoing care.

drag the line — or focus it and use the arrow keys

to my petal's homepage: 'Write your plants a love letter.' beside a pink envelope with a handwritten letter to a plant and sticky care notes — the homepage — a letter to Rosa

p. 04 — warm, botanical, adaptable

Ivory, forest green, editorial type.

Warm ivory, forest green, editorial typography and subtle motion make it feel like a journal rather than a diagnostic tool.

Mobile navigation
Keyboard-accessible photo comparisons
Reduced-motion preferences
Installable PWA

ivoryforestpetal

p. 05 — engineering & machine learning

A real backend, a fine-tuned model.

FrontendReact and TypeScript with Vite, Tailwind CSS, React Router, TanStack Query and Motion

FastAPI backendauthenticated sessions, private image storage, SQLAlchemy + SQLite, Alembic migrations

The modelMobileNetV3-Large fine-tuned with PyTorch across 15 classes — tomato, potato and bell pepper

tested with Vitest, Playwright and pytest

p. 06 — the analysis flow

Allowed to say “inconclusive”.

The analysis keeps model predictions, user observations and curated guidance separate. Confidence calibration and input checks let the app return an inconclusive result when the evidence is insufficient.

p. 07 — results

Real garden photos made the difference.

Fine-tuning with real-world photographs, measured on a held-out set of 310 images:

Macro F1

0.28

0.55

Accuracy of findings shown

37%

79%

before after fine-tuning

p. 08 — what I learned

Communicating uncertainty is part of building a useful AI product.

to my petal strengthened my experience across product design, full-stack development and machine learning evaluation.

— with love, E.H.