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

What I Built

My neighbourhood in my hometown has changed a lot in ten years. The lane outside our society used to look huge, and now it is full of cars. There are far fewer green playgrounds, and the hospital near the one where children played cricket every evening has closed. But the small shops at the corner are still exactly where they were.

SmritiPath turns a memory like this into a walking route. You describe a place the way you remember it, a local model picks out each place you mention, and you pin them on a map. The result is a Walk Pack: one page with a sketch map, what each stop was like then, questions to ask the person walking with you, and a blank "Now" line to fill in by pen. The AI does its work before the walk. Then the screen goes away and the walk is just you and someone who remembers.

It is for anyone who has a person in their life who remembers how a place used to be: a parent, a grandparent, a neighbour. The goal is to get the two of you outside, standing where the memory happened.

Demo

A short screen recording of the full flow:

  1. Paste a memory of a place (or upload a recording) and click Find the stops.
  2. Each stop shows a quote from the text, checked by code.
  3. Pin each stop on the map and reorder the route.
  4. Download the Walk Pack, which works offline.

Code

GitHub logo vaidehi-snk / smriti-path

SmritiPath

Turn a spoken or written memory of a place into a walking route you take together. Speech-to-text and stop-finding run locally with open-source models. Nothing is sent to any AI service.

How it works

  1. Add a recording (faster-whisper) or paste a memory as text.
  2. A local model (gemma3:4b via Ollama) finds each place mentioned, with a quote from the speaker. Code checks every quote really appears in the transcript.
  3. Pin the stops on a map (Leaflet + OpenStreetMap).
  4. Download the Walk Pack: one offline page with a sketch map, "then" stories, questions to ask, and blank "Now" lines to fill in by pen on the walk.

Run it

pip install -r requirements.txt ollama pull gemma3:4b uvicorn server:app

Open http://127.0.0.1:8000

Audio and transcripts are git-ignored and never leave your machine. Map tiles and place search need internet while building; the Walk Pack itself works offline.


View on GitHub


How I Built It

  • Speech to text: faster-whisper (small model, CPU)
  • Finding the stops: gemma3:4b through Ollama, with JSON-schema output
  • Server and page: FastAPI, Leaflet and OpenStreetMap
  • Walk Pack: a single HTML file with an SVG sketch map and no network requests

My first version had problems. It named stops "tree" and "cart", added a stray "quote:" label to quotes, and merged two places into one sentence the speaker never said. I fixed it with a stricter prompt and a worked example, and by adding a check in code that confirms every quote really appears in the transcript. Anything unverified is flagged on the page. A 4B model is not perfect, so every stop is editable before the Walk Pack is built.

Limits: I have tested speech-to-text in English so far, and Hindi and Marathi are next. The map and place search need internet while building, but the Walk Pack itself works offline.

Why Does Open Innovation Matter?

Family memories are personal. With a closed cloud API I would have to upload my family's stories and voices to someone else's server. Here they never leave my laptop, it costs nothing to run, and the model is one line in the code, so swapping it for another is easy.

What's Next

  • Test speech-to-text properly in Hindi and Marathi, since many family memories are told in those languages.
  • Compare a few different open models for finding stops and see which handles messy, spoken memories best.
  • Add an audio version of the Walk Pack, so the questions can be listened to instead of read on a screen.