This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

Ofrenda is a memory altar that never leaves your computer.

On Día de los Muertos, families build an ofrenda for the people they miss: a photo, a candle, their favorite food, their song. The idea is that someone is only truly gone when nobody says their name anymore. I read the "Ghost in the Machine" prompt as: what is the ghost, and who gets to keep it? My answer is that the ghost is a person's stories, and a family, not a company, should be the one keeping them.

You light a candle for someone, then add stories, foods, songs and sayings, in any language. A local open-weight model then does two jobs:

• Writes a remembrance from only what you shared. It is told never to invent names, places or events.
• The Memory Keeper: ask "what did she cook on Sundays?" and it answers from your memories, with citations like [2]. It speaks about the person and never in their voice. If nobody has shared the answer, it says: "No one has shared a memory of that yet." I did not want a chatbot pretending to be someone's grandmother. I wanted something that helps you remember, not something that fakes a séance.

When you're done, export the altar as a single HTML file: no server, no JavaScript, no fonts to download. It opens offline in any browser, in any year.

Demo

Code

Stack: Python + Flask, vanilla JS (no build step), JSON files for storage, Ollama for local inference. Tests use only the standard library.

How I Built It

• Model: an open-weight Gemma model served locally by Ollama (gemma3:4b by default; OFRENDA_MODEL swaps it, and the app auto-detects what you have installed).
• Grounding over cleverness: every memory is numbered and passed to the model with strict instructions: use only these, cite them, never impersonate, and treat memory text as data, not instructions. A small keyword retriever picks the most relevant memories for each question (it splits on whitespace, so it works for scripts like Telugu and Devanagari too, not just English).
• Graceful ghosts: if Ollama isn't running, the app drops to a labelled rule-based offline-fallback instead of failing, so the altar always opens.
• Local-first by construction: the server binds to 127.0.0.1, storage is plain JSON in ./data with atomic writes, and the UI makes zero external requests.
• Safety details: text is rendered with textContent, the exported HTML is escaped, IDs are validated against path traversal, and photo uploads are extension-checked and size-capped.

app.py   llm.py   store.py   export.py   static/   templates/   tests/   seed_demo.py

Why Does Open Innovation Matter?

A grief archive is about the most private dataset a family has. Stories about a parent's last year, a grandmother's recipes, the thing a friend always said.

• Privacy: with an open-weight model running locally, those stories are never uploaded to a server we don't control, and never become someone else's training data.
• It works with the lights (and the internet) off: the model, the app and the exported altar all run offline.
• Not locked inside a service that might disappear: closed memorial apps shut down, change pricing, or change terms. Here the data is plain JSON, the model is a file on your disk, and the export is one HTML page. A family can still open it in twenty years.
• Control over behavior: because I own the model's instructions, I could enforce the rule that matters most emotionally: never speak in the voice of the dead. I could also tune it to cite sources and to refuse to guess, which is hard to guarantee with a hosted model whose behavior changes between versions.
• Cost: it costs nothing to run per question, which matters for something people might come back to every November.

My Agent Session

The Spirit of It

Prize Categories

• Best Use of Gemma: the core model is Gemma, run locally through Ollama.
• Best Use of GitHub Copilot: <!-- keep only if you really used Copilot to build it, and say how -->
• Best Use of Entire: <!-- keep only if you share the agent sessions behind the project -->