Sanpo 散歩: I built a local Japanese tutor from my phone gallery at midnight — no cloud, no walk needed

Salugu Harshita Bhanu — Mon, 12 Oct 2026 06:38:19 +0000

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

What I Built

I'm in my final year of B.Tech. Next year I move to Japan for my first job, and somewhere between the offer letter and the packing list I realised I'm about to leave behind the places that made me — Vizag's beaches, Araku valley mornings, the smell of home.

Sanpo (散歩, "a stroll") is a walking companion that runs entirely on my laptop. You take photos — from a real walk, or from your phone gallery at midnight — and a local Gemma 3 vision model turns what it sees into Japanese flashcards: the word, the hiragana reading, romaji, an example sentence, and a note on how that thing might feel different in Japan. At the end, it writes a short postcard from you to yourself, sealed until your first week in Japan.

Everything — the AI, your photos, your letters — stays on your device. No cloud. No API key. No account.

There's also a weather-based walk quest (Open-Meteo, no key needed), and a personal bucket list of places I want to see before I leave India — RK Beach, Araku Valley, Borra Caves, Dudhsagar Falls, Darjeeling...

I couldn't go outside — it was midnight. So I used what I had: a rainy windshield from a monsoon drive, my room with Diwali string lights and a suitcase in the corner, a yellow rose, an idli-dosa thali, my boots, and a piece of art I love. The model found words in all of them.

Demo

Code

Git-brintsi20 / Sanpo

A bit of something Before I leave here and begin Again .

Sanpo 散歩 — Before I Go

A local-first walking companion for learning Japanese through what you see.

I'm a final-year student in Bhopal, moving to Japan next year. This app turns a walk into Japanese flashcards and a postcard to my future self — all running 100% on my laptop, no internet required after setup.

What it does

Photo → Japanese Cards — Upload walk photos; a local Gemma vision model picks 3 objects and returns Japanese word cards with hiragana, romaji, an example sentence, and a note on how the thing might feel in Japan. Hit 🔊 to hear the pronunciation.
Postcard to Future Me — At the end of a walk, write a short letter to yourself in your first week in Japan. Seal it with a date; it shows as an envelope with a countdown until then.
Weather Quest — Fetches Bhopal's current weather (Open-Meteo, no key)…

View on GitHub

How I Built It

Model: Gemma 3 4B (vision), served locally by Ollama. Everything runs on CPU — 32 GB RAM, no GPU.
App: Next.js 14 (App Router) + TypeScript + Tailwind CSS. Three API routes call Ollama at localhost:11434 — vision analysis, postcard writing, and weather quests.
Speed: On CPU, gemma3:4b took between 10.4s and 65s per photo. Simple images (boots: 10.4s, food: 13.7s) were fast. Complex ones (room with Diwali lights: 65.3s) took longer. Client-side resize to 768px before sending was the biggest single speed win.
Extras: Open-Meteo for weather (no key), Web Speech API for ja-JP pronunciation, localStorage for all walks and postcards. Zero external dependencies for AI.
12B model: Didn't have disk space for gemma3:12b (~8 GB needed). Ran everything on 4B — for everyday objects, it was more than enough.
What actually went wrong: Two real mistakes. A rainy windshield photo returned あられ (arare), labeled "Raindrops" — but あられ means hailstones. Raindrops are 雨粒 (amatsubu). And a yellow food plate returned ヒマワリ (himawari) labeled "Marigold" — himawari means sunflower. The model saw yellow, guessed flower, got both the word and the English wrong. I kept them in. That's honest.

The postcard Gemma wrote me (sealed until 12 October 2027):

"It's strange, isn't it? This morning walk felt so familiar, even though I'm so far from Vizag. I spent a while just appreciating the bara and ha — the cherry blossoms and leaves. I learned words like arare, Iruminēshon, and Suitukēsu. I'm feeling a little quiet and wishful, a little homesick perhaps. It's a lot, but I'm trying to be patient. — You, today"

It hallucinated a lake. But the homesick part it got exactly right.

Why Does Open Innovation Matter?

A postcard to my future self is about as personal as data gets. With an open-weight model running locally, my photos and letters never leave my laptop. I don't pay per token for something I want to reread in a year.

It also runs on weak hardware with no internet after setup. The model is honest about being small — it mistakes hailstones for raindrops — and somehow that feels more trustworthy than a polished API that hides its errors.

These are the memories that will keep me company when I feel lonely in Japan one day, missing home. Building something that holds them felt right.

Prize Categories

Best Use of Gemma

DEV Community: Salugu Harshita Bhanu