Ofrenda: An Offline Family Memorial with Local AI

"What I Built" - Ofrenda is a private, offline digital altar designed for families to remember those they have lost, particularly during Día de los Muertos. This concept extends beyond traditional ofrendas to include festivals, rituals, and recipes that symbolize a family's enduring legacy. The altar accommodates a person, a festival, or a ritual, effectively capturing the essence of what a family carries forward.

Users can contribute to the altar by recording a story, uploading an audio or video file, writing a narrative, or sharing a photo. Each memory can be customized with its own photo and is categorized as a Recipe, Ritual, Festival, or Story. Both audio and video files are transcribed locally using the "faster-whisper" speech-to-text model powered by the Whisper small model. Non-English speech is translated into English, and the app keeps the original transcript alongside an English version for reference.

The creation of a remembrance is a simple one-click process, which compiles favorite foods, songs, and sayings to honor the remembered individual or event. The project’s underlying philosophy is encapsulated in the statement, "the memories we carry with us," emphasizing that no memory is ever truly lost.

The technical backbone of Ofrenda consists of locally run models: "faster-whisper" for speech recognition, Llama 3.2 3B running through Ollama for translation and text generation, along with a backend built using FastAPI with Uvicorn and SQLite for storing data. The frontend is constructed from plain HTML, CSS, and JavaScript without any build process, ensuring a seamless, offline user experience.

Open innovation plays a crucial role in this project by enabling families to maintain the privacy of their personal memories. Since the app operates without internet connectivity, users can run it on their laptops without the need for accounts, API keys, or cloud dependencies. This offline capability ensures that all data, including voice notes, videos, photos, and stories, remains stored locally, safeguarding against any potential service disruptions or shutdowns that could otherwise compromise the family’s digital legacy.

The project is not only cost-effective but also resilient, as it eliminates reliance on subscription services that could disappear overnight. Additionally, the models used in the app are highly customizable. Families have the flexibility to fine-tune the remembrance style based on personal letters or switch to models better suited to their language, allowing for a more tailored and accurate remembrance experience.

However, it’s important to note that small local models may occasionally produce errors, particularly with names, festival terminology, and translations. To address this, the app displays the original transcript and its translated version side by side, prompting families to review and verify each remembrance before finalizing it.