Ofrenda: An Offline Family Memorial with Local AI

What I Built

Ofrenda is a private, offline digital altar for families. For Día de los Muertos, families set up an ofrenda to remember the people they've lost. I took that idea further: the altar can hold a person, a festival, or a ritual, because those are also what a family carries forward. Ganpati, Diwali lamp walks, a grandmother's recipe, and her voice telling a story all belong on the altar.

You can record a story in the browser, upload an audio file or a video, write a story, or upload a photo. Each memory can carry its own photo and is marked as a Recipe, Ritual, Festival, or Story. Audio and video are transcribed locally. When a story isn't in English, the app keeps the original transcript and adds an English version. Then one click creates a remembrance with favorite foods, songs, and sayings.

I read the theme as "the memories we carry with us." Nothing is ever lost, because the stories are kept on the family's own machine.

Code

https://github.com/Nivi61/Ofrenda

How I Built It

• Speech-to-text: faster-whisper with the Whisper small model. Its translate mode turns non-English speech into English.
• Remembrance writing: Llama 3.2 3B running locally through Ollama.
• Backend: FastAPI with Uvicorn, and SQLite for storage.
• Frontend: plain HTML, CSS, and JavaScript, with no build step.

Every model runs on the laptop. The app has no cloud dependencies, so nothing needs an account or an API key.

Why Does Open Innovation Matter?

This project is about family memories, which are private. Open models make it possible to keep them private:

• It works with the Wi-Fi off. After the models are downloaded, the whole app runs offline, which suits a memorial.
• The family's data never leaves the laptop. Voice notes, videos, photos, and stories stay in a local folder.
• It costs nothing to run, and it doesn't depend on a subscription or a service that could shut down and take the family's stories with it.
• The models can be swapped or tuned. A family could later fine-tune the remembrance style on their own letters, or switch to a model that works better in their language.

Honest Limitations

Small local models can make mistakes, especially with names, festival terms, and translations. The app shows the transcript and translation next to each other, and the family should check the remembrance before keeping it.