🕯️ Ofrenda

On Día de los Muertos, families build an ofrenda for the people they miss: a photo, a candle, their favorite food, their song. Someone is only truly gone when nobody says their name anymore.

A family's grief archive is some of the most private data they own, so it should not live on a server they don't control. Ofrenda keeps everything on your computer and uses an open-weight model running locally to help you remember.

• 🕯️ Light a candle: create an altar with a name, years, relationship and photo.
• 📖 Add memories in any language: stories, favorite foods, songs, sayings.
• ✍️ Write a remembrance: the local model drafts a short tribute using only your memories. It is told never to invent names, places or events. You can edit it before saving.
• 🗣️ Ask the Memory Keeper: ask "What did she cook on Sundays?" and get an answer with citations like [2].
  • It speaks about the person, never in their voice. No fake séance.
  • If nobody shared the answer, it says so instead of guessing.
• 📦 Export as one HTML file: no server, no JavaScript, no external fonts. Opens offline in any browser, in any year.
• 🛟 Graceful fallback: if Ollama isn't running, a labelled rule-based mode keeps the app usable.

Install Ollama, then pull a Gemma model:

ollama pull gemma3:4b
ollama list        # confirm the exact model name you have

Windows (PowerShell / VS Code terminal)

git clone https://github.com/noorulhuda07/ofrenda.git
cd ofrenda
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python seed_demo.py     # optional: creates a fictional demo altar
python app.py

macOS / Linux

git clone https://github.com/noorulhuda07/ofrenda.git
cd ofrenda
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python seed_demo.py
python app.py

Open http://127.0.0.1:5000.

The first model request loads the model into memory and can take a minute on a CPU-only laptop. Try a smaller model such as gemma3:1b if it's too slow.

PowerShell example: $env:OFRENDA_MODEL="gemma3:1b"; python app.py

Browser (vanilla JS)  ->  Flask API (127.0.0.1)  ->  Ollama (local Gemma)
                               |
                          ./data/*.json  (plain files you own)

1. Memories are numbered and passed to the model: [1] (food) tamales de rajas ...
2. Strict system prompts tell the model to use only those memories, cite them, never impersonate the person, and treat memory text as data rather than instructions.
3. A small keyword retriever selects the most relevant memories per question. It splits on whitespace, so it works for scripts like Telugu and Devanagari as well as English.
4. Citations are parsed from the answer and shown next to it so you can see exactly where each claim came from.
5. Export renders the altar into a single escaped HTML file with the photo embedded.

app.py            Flask server + JSON API
llm.py            Ollama client, prompts, retriever, offline fallback
store.py          JSON-file storage (atomic writes, validated ids)
export.py         One-file offline HTML export
templates/        index.html
static/           app.js, style.css (no build step, no external requests)
seed_demo.py      Fictional demo data for screenshots and demos
tests/            Unit tests (standard library only)

• Server binds to 127.0.0.1 only. Nothing is sent to any cloud service.
• Data is plain JSON in ./data: readable, copyable, deletable. data/ofrendas/ and data/photos/ are git-ignored, so real memories are never committed by accident.
• The UI renders user text with textContent, and the export HTML-escapes everything.
• IDs are validated against path traversal; photo uploads are extension-checked and capped at 8 MB.
• Open-weight models can still make mistakes. Treat drafts as a starting point and edit them.

python -m unittest discover -s tests -v

Covers storage validation, retrieval (including Telugu tokens), offline fallback, and HTML-escaping in the export.

• Private: the stories never leave your machine or train someone else's model.
• Offline: the model, the app and the exported altar all work without internet.
• Durable: not tied to a memorial service that can change terms or shut down.
• Controllable: owning the prompts means I could enforce the rule that matters most, never speak in the voice of the dead.
• Free to run: no per-question cost.

• Voice-note recording with local Whisper transcription (oral history)
• Printable ofrenda poster export
• Multiple photos per altar
• Local embeddings for better retrieval on large archives

Issues and PRs are welcome. Please never include real people's private memories in issues, screenshots or tests; use fictional data like seed_demo.py.

MIT © 2026 Noor Ul Huda

Ollama, Gemma, Flask, and the DEV + MLH Hacktoberfest community.