What I Built

I built LOOK AGAIN, an app that doesn't tell you what's around you. It finds out what you missed.

The idea started with a weird thought: we walk through the world all the time, but we remember almost none of it. Someone could be sitting 15 minutes behind us and we'd never know. So instead of another "identify this bird" app, I made one that trains attention and then checks your memory.

Here's the loop:

Pack (at home): You pick your time, setting, and focus (sounds, above eye level, small details, and so on). A local model writes a short "look" prompt plus hidden check questions, like "What's the highest man-made thing in view?" or "What's directly behind you?"

Look (screen off): Put your phone away for 5, 10, or 15 minutes and just notice.

Recall: Back turned, answer the questions from memory.

Look Again: Turn around, check your answers against reality, and mark each Right, Wrong, or Never noticed. No camera, no vision model. The real world is the answer key.

Journal: Everything is saved in a private field journal on your device, and you can export it as a readable text file.

To make people come back, I added two things that only work because the AI runs locally:

Field Stories: After each completed walk, the model writes the next chapter of a short mystery where your real observations become the clues. A new chapter only unlocks when you take another walk. It's clearly labeled as fiction.

Echo: The next day, the app asks what you still remember about yesterday's walk, then shows what stuck, what faded, and what only came back now.

It's for anyone who wants to look up from their phone, especially people with no trails nearby. There are no points, streaks, or leaderboards on purpose.

Code

vinithalakhwani6869-ops / Look-Again

LOOK AGAIN — Open-Source AI Outdoor Attention Companion

"There's more around you than you notice."
Built for the Touch Grass hackathon theme: open-weight AI that gets people outdoors, where the screen is the shortest part of the experience.

1. What is LOOK AGAIN?

Most people move through their everyday neighborhoods, parks, and campuses without remembering much of what was right in front of them. LOOK AGAIN makes the gap visible: what you were sure you saw versus what was actually there.

Phone away first: You get one single, clear observation focus for a 5, 10, or 15-minute walk. Then your phone goes into your pocket.
Memory meets reality: When the calm timer finishes, you answer hidden check questions strictly from memory with your back turned. Then you turn around and physically look again. You are the judge.
Zero surveillance or scoreboards: No cameras, no computer vision, no species identification quizzes…

View on GitHub

How I Built It

Model: Qwen3 4B (Apache 2.0), an open-weight model run locally with Ollama.

Frontend: React + Vite + Tailwind, with a cinematic Home screen and a plain black-and-white theme for the rest so the screen stays quiet.

Backend: A small Express server that talks to Ollama, uses structured JSON schemas so the small model returns valid output, validates everything it sends back, and queues requests so only one generation runs at a time.

Storage: The journal lives in the browser (localStorage). There's no account and no database.

Fallbacks: If Ollama isn't available, the app says so and uses clearly labeled preset packs, so it never fakes AI output.

I built this on a CPU-only laptop, and generating a pack can take up to a minute or so. That shaped the whole design: you prepare the walk before you leave, so the AI isn't needed while you're outside.

Why Does Open Innovation Matter?

A field journal is full of raw, unedited memories. With a closed API, every one of those would have gone to someone else's server. With an open-weight model running on my own machine, the model that writes your stories and sorts your Echo never sees a cloud service.

Open models also gave me things I couldn't get from an API:

It costs nothing to run. No keys, no usage bills.

I can swap the model by changing one line in a config file.

I can see exactly what the model is allowed to do, so I can validate its output and keep it from inventing observations you never made.

One honest note: The live demo link above runs in preset mode only. It has no AI model behind it, so packs come from a built-in library, and Field Stories and Echo sorting use their fallbacks. The real open-source AI (Qwen3 4B via Ollama) runs on your own machine. To try the full version, follow the 3-step setup in the GitHub README. The demo video shows the complete experience with the local AI running.

DEV Community

Look Again