An AI field guide that wants you to put your phone away

Quick start

Use Chrome or Edge on a laptop or desktop: they support WebGPU, which makes Gemma fast. Other browsers may work but haven't been tested.

1. Download the models once

   Open Setup (or the "Get started" card) and click Download models. It's about 900 MB, so use Wi-Fi. This happens only once, and the app stays usable while it downloads.

2. Plan a walk

   On Go outside, pick a time and a place and press Plan my walk. You get 3–5 things to look for in about ten seconds.

3. Walk, spot, identify

   Press Start walking. The screen turns into a timer and checklist. When something catches your eye, tap I found something, take one photo and save it to your journal.

Going for a walk

The walk planner is the heart of the app. It gives you a reason to go out and things to notice once you're there.

How long?

15 minutes gives you 3 things to find, 30 minutes gives 4, and an hour gives 5.

Where?

Streets, park, woods, garden or water. The list changes to match: city walks get cracks in pavement and rooftop birds, woods get bark, moss and tracks.

Season

Worked out from today's date, so autumn lists include falling leaves and seeds. If you've saved a location south of the equator, the seasons flip.

Variety

Things you've already saved to your journal are left out of new lists. Don't like a list? Press Give me different ones.

Walking mode

Once you press Start walking, the app hides everything except the timer and your checklist. That's on purpose: the less there is to look at, the more you look around.

• Tick items off as you find them. The counter shows how many are left.
• The timer can run over. It counts up past zero in green. Nobody is rushing you home.
• Closed the tab by accident? Your walk resumes where you left off.
• End walk saves your minutes outside and how many you found to the journal stats.

Identifying a find

Use Identify or the I found something button during a walk. You can drop a photo in, paste one (Ctrl+V), choose a file, or use your webcam.

Taking a photo that works

Reading the result

You get a name, a group (bird, tree, flower, bug, mushroom...) and a confidence bar.

Confident

60% or more. Usually right for common species. Still use the field note to check.

Likely

30–60%. Look at the other options too: similar species often score close together.

Unsure

Below 30%. Either it isn't in the field guide yet or the photo is hard. Try a closer, clearer shot.

Not right? Tap one of the other options under the name. The field note rewrites itself for that species, and your choice is saved as "you confirmed".

Reading a field note

After every identification, Gemma writes a short note. It isn't an encyclopedia entry. It's a nudge to look more closely and keep exploring.

Check it

Two features to confirm in person. This is how you catch a wrong ID: if the real thing doesn't match, pick another option.

Look nearby

One related thing to go and find next, which keeps the walk going.

Careful

A safety or respect tip. These come from a fixed, hand-written list, never from the AI.

Why you can trust it more than a chatbot

Small AI models sound confident even when they're wrong. So Gemma is never asked to remember facts. Each species in the field guide has a line of hand-written ID traits, and Gemma only rephrases those into things to check. Every note is then checked automatically. If it leaves out a section, drifts from the trusted traits or repeats the instructions, the app replaces it with a plain note built straight from the traits.

Your journal

Every find you save goes into Journal with its photo (as a small thumbnail), the time, your own note and the field note. The stats at the top count your minutes outside, walks, finds and different species.

• Location is opt-in. Tick "Save the location with this find" and the browser asks for permission once. Coordinates are rounded to about 10 m and stored only on this device.
• Export as Markdown downloads your whole journal as a text file you can keep, print or paste into a notes app.
• Delete removes a find for good.
• The journal lives in this browser. Clearing site data, or using a private window, erases it, so export it now and then.

Using it offline

After the first download, the app, both models and your journal are all stored on your device.

Check it really works offline

1. Open Setup and wait until both models say Ready, saved for offline use.
2. The status pill at the top should read Ready, works offline.
3. Turn on airplane mode (or disconnect Wi-Fi) and reload the page. It should load and the pill should read Offline and ready.
4. Identify a photo. If it works, you're set for the trail.

What the status pill means

Models not downloaded

Nothing is downloaded yet. Open Setup.

Loading models…

Downloading, or loading from your device (about 15–20 s on later visits).

Ready, works offline

Everything is stored locally.

Ready (online only)

The models work but couldn't be saved, usually because the browser is low on storage. See troubleshooting.

Settings

Model size

1B (default, ~785 MB) writes clearly better notes and lists. 270M (~300 MB) is smaller and faster, but its writing is weaker, so more notes fall back to the plain template.

Run on: Auto

Uses your graphics chip (WebGPU) when the browser supports it. If that fails, it switches to the CPU by itself and remembers that for next time.

Run on: GPU / CPU

Forces one or the other. The CPU always uses the 270M model, because the 4-bit 1B model can't run on the browser's CPU backend.

Remove models

Deletes the downloaded models to free space. Your journal is kept.

Typical speed on a mid-range laptop (Intel Iris Xe): identifying a photo takes 1–2 s, and a field note takes about 10 s on the GPU.

Troubleshooting

Privacy

Nothing you do in the app leaves your device: no photos, notes, locations or journal entries. There is no account, analytics or server.

• The only network requests are loading the app itself and downloading the model files from Hugging Face the first time.
• You can check this yourself: open your browser's developer tools, go to the Network tab, and identify a photo. Nothing is sent.
• Data is stored in the browser's own storage for this site (IndexedDB for the journal, the Origin Private File System for models). Removing the site's data in your browser settings erases all of it.

Add your own species

The field guide is a plain list in src/labels.js. Each entry has a name, a group and a line of ID traits:

['hoopoe', 'bird', 'pinkish-orange body; black-and-white striped wings; fan-like crest tipped black; long curved bill'],

1. Add your local species with traits a beginner can check by eye.
2. Run npm run embed to update the field guide's embeddings (a minute or two).
3. Run npm run build and redeploy, or use npm run dev locally.

No training or fine-tuning: SigLIP recognises the new names right away, and Gemma uses your traits in its notes.

How it works

Built with Transformers.js and ONNX Runtime Web. Models: Gemma 3 1B (Gemma Terms of Use), SigLIP base (Apache-2.0). Source code is MIT-licensed.