This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
What I Built
Siempre ("Some stories deserve to stay") is a private, AI-powered digital memorial and living ofrenda inspired by DĂa de los Muertos and personal remembrance traditions.
The Interpretation of "Ghost in the Machine"
As Halloween and DĂa de los Muertos approach on November 1st and 2nd, tradition holds that the boundary between past and present grows thin. But our digital tools often treat memory with cold sterility—a forgotten folder of JPEGs in the cloud or a detached chat interface.
Instead of building another hollow chatbot pretending to channel the dead, I built a tactile, warm, and reverent digital altar where family memories, authentic voices, favorite foods, and heirloom objects remain alive. Every object on the altar has a true story behind it:
- Tiered Interactive Altar: A customizable 3-tiered ofrenda adorned with handcrafted golden cempasĂşchil (marigolds), flickering candlelight vigils, papel picado banners, clay cazuelas of mole poblano, sweet pan de muerto, and vintage polaroids.
- Family Voice Archive: An in-browser audio recorder and player with interactive timestamped transcripts. Clicking any transcript snippet jumps playback directly to that moment in the original recording.
- Ethical AI Remembrance Studio: A local-first narrative weaver powered by open-weight models (Ollama / Llama 3.2 / Gemma). It drafts celebratory, warm tributes grounded strictly in user-entered memories—with hard constraints forbidding fabricated quotes, fake stories, or pretending to speak for deceased loved ones.
- Archive Search & Rediscovery: Instant keyword search across all altar items, spoken audio transcripts, and memory slips with context highlighting.
-
Complete Preservability: One-click JSON archive export and print-ready keepsakes (
@media print) so family memories are never trapped in a proprietary walled garden.
Demo
đź”— Live Application: https://siempre-mauve.vercel.app/
(Tip: You can also test the full interactive experience locally with sample data by running ./start.sh or exploring the pre-seeded "Abuela Elena Rosales" demonstration memorial.)
Code
Trideep091
/
siempre
A Living Digital Ofrenda Powered by Open-Source AI
Siempre — A Living Digital Ofrenda
"Some stories deserve to stay."
A little place for the people who will always be with us.
Siempre is a private, AI-powered digital memorial and living ofrenda inspired by DĂa de los Muertos traditions. It preserves family memories, photographs, spoken stories, heirloom objects, favorite foods, songs, and remembrance notes.
Built with an artisanal, scrapbook-inspired visual design, authentic SVG artwork, and a local-first open-source AI architecture that guarantees personal memories stay securely on your computer.
Key Features
-
The Interactive Ofrenda (Altar Workspace):
- Layered tiered composition: Top Tier (portraits, sacred space, vigils), Middle Tier (favorite dishes, songs, keepsakes), Base Tier (golden cempasĂşchil marigolds, written memory slips).
- Dynamic item cards: Polaroid-framed photographs with rotation, Mexican clay cazuela bowls (mole poblano), sweet pan de muerto bread, music cards with direct play links, flickering wax candles.
- Click any object on the altar to open its story modal, recipe notes…
Repository: https://github.com/Trideep091/siempre
- Frontend: React 18, TypeScript, Vite, Tailwind CSS, bespoke SVG artwork, and canvas-confetti.
- Backend: FastAPI, SQLAlchemy (asyncio), SQLite / aiosqlite, Pydantic v2.
-
Deployment: Vercel Multi-Service architecture (
vercel.json), single-command production deployment script (deploy.sh), and multi-stage Docker container (Dockerfile).
How I Built It
Siempre is engineered entirely around open-source AI, local-first inference, and ethical guardrails:
-
Open-Weight Language Models via Local Inference (Ollama):
- The Remembrance Studio interfaces directly with open-weight models (like
llama3.2:3b,gemma2:2b, ormistral:7b) via local Ollama inference (http://localhost:11434). - The prompt architecture enforces strict grounding transparency: users select exactly which altar keepsakes, favorite dishes, and audio stories the model may reference. If a fact isn't in the provided notes, the model is instructed never to invent it.
- The Remembrance Studio interfaces directly with open-weight models (like
-
Open-Source Speech Recognition (
faster-whisper):- Spoken family stories recorded via the browser MediaRecorder API are transcribed locally using open-source Whisper models.
- Preserves timestamps for every sentence so listeners can scrub through family recordings interactively.
-
Graceful Cloud Fallback:
- In cloud environments where local hardware inference isn't running (e.g. on serverless Vercel), the application gracefully reports local AI engine status and allows full manual editing, transcript correction, and audio archiving without breaking or throwing errors.
Why Does Open Innovation Matter?
Preserving the memory of someone you love is an inherently sacred, vulnerable human act. Closed-source, commercial AI clouds fundamentally fail this use case:
Privacy & Grief Are Not Training Data:
When someone records their grandmother’s voice or uploads personal photos and handwritten notes, they should never have to wonder whether their grief is being scraped to train a commercial model or monitored by third-party telemetry. Open-source models running locally on the user's computer guarantee that intimate family memories never leave the room.The "Forever" Test (Longevity):
Proprietary APIs shut down, change pricing, deprecate endpoints, or pivot business models. An open-weight model (llama3.2,gemma,whisper) running in an open-source container will still run 10, 20, or 50 years from now. A memory space must be durable enough to outlive cloud startups.Ethical Guardrails Over Sycophancy:
Commercial consumer chatbots are optimized for user retention and engagement, often hallucinating flattering fiction or pretending to channel the deceased. Open innovation allowed me to inspect, control, and strictly enforce the prompt boundaries: AI assists in polishing the prose of the living, but never puts fabricated words into the mouths of the dead.
My Agent Session
This project was designed, implemented, and deployed with the assistance of Google Antigravity, autonomously scaffolding the full-stack architecture, designing original SVGs, implementing local speech and LLM pipelines, and configuring Vercel multi-service orchestration.
The Spirit of It
When building the demonstration ofrenda, I created Abuela Elena Rosales—a grandmother from Pátzcuaro, Michoacán.
On her altar, you find a steaming clay cazuela of mole poblano, sweet pan de muerto, a vintage polaroid of her courtyard garden, and an audio memo recorded by her grandson Carlos recalling how she used to slip lemon drops into small eager hands while baking.
When you click on her voice recording and hear the gentle courtyard sounds while reading her story, it doesn't feel like an "AI feature"—it feels like opening a cherished wooden keepsake box tucked away in an attic. That is what AI should do: honor what is real, rather than fabricating what is not.
Prize Categories
- Best Use of Gemma: (Configured and tested with open-weight Gemma models via local Ollama inference)
- Overall Winner: Hacktoberfest Open-Source AI Challenge: Week 4
Top comments (0)