Echoes of the Ofrenda

Project Name: Echoes of the Ofrenda (Ecos de la Ofrenda) 

Project Overview
Echoes of the Ofrenda is a fully local, privacy-first oral history and digital ofrenda builder designed to honor ancestors for Día de los Muertos. The application runs entirely on a laptop without an internet connection. It records spoken family stories, transcribes them using an open-source speech-to-text model, and utilizes a local open-weight Large Language Model (LLM) to extract memories, favorite foods, songs, and distinct traits of departed loved ones. It then automatically structures these into a beautiful, personalized digital remembrance page and generates a custom Calaverita (a traditional humorous or affectionate poem written for Day of the Dead). 

How It Works 

1. Voice Capture: A family member speaks into the mic, sharing unscripted memories, stories, or anecdotes about a loved one.
2. Local Processing: The audio is transcribed locally using OpenAI's open-source Whisper-base model via the faster-whisper library.
3. Ghost in the Machine Engine: The text is processed by a local Llama-3-8B-Instruct model (quantized to 4-bit to run smoothly on consumer hardware) via the Llama.cpp framework.
4. Asset Generation: The local LLM strips out conversational filler and extracts key cultural elements (e.g., "He loved pan de muerto and always listened to Mariachi music"). It also writes a custom, rhyming Calaverita based on those specific details.
5. The Digital Ofrenda: The system outputs a clean, markdown-based digital altar that can be preserved forever as a flat file.

Why Open Innovation Matters for This Project
When dealing with family histories, ancestral memories, and deeply personal grief, closed-source AI falls short for several critical reasons: 

Absolute Privacy for Sacred Memories
Family stories, old photos, and vulnerable reflections about the deceased should not live on a corporate server, be used as training data, or be subject to data breaches. By running inference entirely locally, users have the absolute certainty that their family's private oral history remains strictly on their own machine. 

Works with the Lights (and Internet) Off
True to the theme, this machine works when the world goes dark. Whether you are in a rural area visiting family gravestones with zero cellular reception, or experiencing a power outage, the entire pipeline operates smoothly offline. The spirits are channeled entirely from local silicon. 

Immortal Software for Imperishable Memories
Closed-source APIs change, deprecate, or disappear behind shifting paywalls. If this project relied on a commercial LLM API, the digital ofrendas built today could break down in a few years when that specific endpoint is retired. By anchoring the project to open-weight models and open-source frameworks, the tool belongs entirely to the creator. It can run 10, 20, or 50 years from now, ensuring these digital altars are as lasting as the memories they preserve. 

Where Open-Source Excelled Over Closed-Source
Closed-source models often carry heavy guardrails or generic alignment filters that tend to flag or sanitize discussions surrounding death, ghosts, or the afterlife as "sensitive content" or "depressing." Local open-weight models allowed us to fine-tune the system prompts specifically for Mexican folklore, grief processing, and traditional Day of the Dead lyricism without triggering corporate safety filters. We achieved a highly specific, culturally authentic tone that a generic commercial API would actively resist. 

Built With 

• Framework: Llama.cpp (Local LLM inference orchestration)
• LLM: Llama-3-8B-Instruct (4-bit quantization)
• Speech-to-Text: OpenAI Whisper (via faster-whisper)
• Interface: Streamlit (running on localhost)