This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Hacktoberfest said “touch grass"... so I built an app that literally makes you do it.

Meet BioDex — an Android game where plants are your Pokémon cards. Snap a pic 🌿📸, wait a few seconds, and boom: you’ve got a collectible card with the plant’s name, rarity, and a fun little field note. XP, a daily streak, and a daily field mission push you to go find the next one. The only way to fill the Dex is to leave the house.

The screen is meant to be the shortest part of the walk: one photo, five seconds of waiting, then the phone goes back in your pocket. This Saturday morning I tried it for 30 minutes in my society’s park and collected 8 cards. It took me back to my biology classes, where we learned the scientific and common names of plants and animals. I used to remember so many of them, like Mangifera indica for mango. It was a nice flashback to my school days.

🌱 Features:

• Camera-based plant capture for collectible cards
• Learn scientific names and common names of flora
• Rarity system (S‑rank, rare, common) for added challenge
• Offline gameplay — no internet required (apart from downloading the model on first run)
• Educational yet fun: explore, collect, and build your own flora deck
• Safe and private: no personal data collected or shared

Demo

A short screen recording shows BioDex in action:

🎮Try it

• Join the closed testing group through this Google Group: BioDex Closed Testers
• Or download the signed APK from GitHub: BioDex-v1.1.3.apk

Code

Everything lives in one repo: the Android app, the model export script, and the Gemma note generator.

How I Built It

BioDex is a Kotlin app: Jetpack Compose for the UI, CameraX for the camera, ONNX Runtime for the model. The interesting part is what the model is and where each half of it runs.

BioCLIP, split in half. Only the image encoder runs on the phone. The 4,271 plant names were turned into embeddings ahead of time and ship as a table. At runtime, a photo becomes a vector, which is compared against that table with cosine similarity. The closest match becomes your card.

Gemma at build time. Gemma 4 E4B, running locally through Ollama, wrote a common name and a short field note for all 4,271 plants, once, with a resumable script. No language model runs on the phone. The notes describe the species, not your photo, and they can be wrong.

Shipping releases. I run a GitHub Actions workflow manually and choose whether the change is a patch, a minor or a major release. It bumps the version, then builds a signed AAB for the Play Store and a signed APK for sharing locally, such as the v1.1.3 link above.

🤦Things that went wrong (lol)

Testing on my own phone turned up two problems:

1. Random objects came back as plants. I photographed things that weren’t plants at all, and BioDex still returned the closest plant in the guide. Similarity always finds a nearest neighbour. I fixed it with a minimum cosine-similarity threshold of 0.45, so below that BioDex says it doesn’t recognise the photo.

2. Plain-looking plants give false matches. If the plant shows no distinguishing features, such as leaves and small buds but no recognisable flower, the model can still return a confident-looking wrong species. The threshold doesn’t catch this, because the photo is plant-like enough to pass. The similarity score isn’t a confidence level, so BioDex treats it as a suggestion, never a verdict.

Other rough edges:

• Specimen #0001: the note said small pale flowers, and my photo was full of pink ones.
• Plants only. Animals and fungi aren’t in the guide, and a plant that isn’t listed can only come back as its closest cousin.
• Never use BioDex to decide whether a plant is edible, poisonous or medicinal.

Why Does Open Innovation Matter?

Where closed would have won. A hosted model would probably know far more than 4,271 plants, handle animals and fungi too, and wouldn't need 1.26 GB of my phone. Open cost me storage, RAM and one very long export script instead of money and privacy. For a game whose whole point is walking somewhere without signal, I'd make that trade again.

I also didn't have to invent any of the hard parts. The Imageomics Institute trained and open-sourced the model, Nate Hamilton published a plant list and embeddings I could build on, Google's Gemma wrote the notes, and ONNX Runtime ran the whole thing on a phone. I wired them together and put a card game on top. That's open innovation at hobby scale.

Prize Categories

Best Use of Gemma. Gemma 4 (E4B), running locally through Ollama, wrote the common name and field note for all 4,271 plants. It runs at build time, not on the phone, and the generator is a resumable script in the repo.

Best Use of GitHub Copilot. I used Copilot to generate the GitHub Actions workflow that prepares releases. I choose patch, minor or major, and it versions the app and builds a signed AAB and APK, so each release is one click to publish to the Play Store or share locally.