Ofrenda:Hacktoberfest 開源 AI 挑戰第 4 週作品
dev challenge w4
Ofrenda 是一款以本機運作為先的應用程式,將語音或文字記憶整理成數碼祭壇,並生成保留講述者語氣的追憶文字。它以 whisper.cpp 本機轉錄語音,透過 Ollama 在本機運行 Gemma 擷取人物、地點、食物和歌曲,資料儲存在 SQLite,無需帳戶或雲端服務。
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
What I Built
Ofrenda is a local-first app that helps you build a digital ofrenda for someone you've lost. You record or type a memory, and it transcribes it, pulls out the people, places, foods, and songs, and arranges them on a candlelit altar. Then it writes a short remembrance in the voice of the story you told, not a generic one.
I interpreted "Ghost in the Machine" as the opposite of the creepy version. The ghost isn't a haunting. It's a memory that keeps living inside the software. My grandmother's recipes and stories existed only in my family's heads, and I wanted a tool that could keep them without sending them anywhere.
Demo
🔗 Live demo: <your-link>
🎥 Video walkthrough: <your-link>
Code
ANAND-debug-bit (ATHARV ANAND) · GitHub
Loved by God and hated by whoever IDGAF. ANAND-debug-bit has 25 repositories available. Follow their code on GitHub.
github.com
How I Built It
-
Transcription: Whisper (open-weight) running locally via
whisper.cpp - Language model: Gemma, served locally with Ollama, for extracting entities and writing the remembrance
- Frontend: a small React app that renders the altar (marigolds, candles, photos) on a canvas
- Storage: everything lives in a local SQLite file. No accounts, no cloud.
The pipeline is simple: audio → Whisper → text → Gemma (prompted to return structured JSON of people, foods, songs, and places) → altar items, then Gemma again for the final remembrance. I spent most of my time on the second prompt. The first drafts sounded like a greeting card, so I fed it the speaker's own phrases and told it to keep their rhythm and quirks.
Why Does Open Innovation Matter?
A family's stories are the most private data there is. Sending a recording of my grandmother's voice to a third-party API, where terms can change and the service can disappear, felt wrong for something meant to outlive me.
Open weights made three things possible that a closed API wouldn't:
- It works with the internet off. I tested it on a laptop in airplane mode, which is how I'd want it to work when my family is gathered around it.
- Privacy is structural, not promised. The audio never leaves the machine, so I don't have to ask anyone to trust a policy.
- It can't be taken away. If a hosted model is deprecated, the ofrenda changes or vanishes. With a model I can pin and keep, the remembrance I generate today will be reproducible in ten years.
My Agent Session
<DevRelay link or embed>
The Spirit of It
I tested it with a story my mother told about my grandfather burning rotis every Sunday and insisting they were "extra crispy on purpose." The generated remembrance kept that line, and she went quiet, then laughed, then asked me to play it again. That was the moment I knew it worked. It didn't invent anything. It just kept the story from getting lost.
Prize Categories
- Google Gemma
來源:DEV Community · dev.to