Ghost in the Machine：以開放權重 AI 驅動的隔網祭壇與附身終端機

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

What I Built

Ghost in the Machine: Anima & Ofrenda is an air-gapped, open-weight AI experience created for the threshold between Halloween (October 31) and Día de los Muertos (November 1 & 2)—a stretch of the calendar when the veil between past and present, living silicon and departed memory, grows exceptionally thin.

Instead of generic horror tropes, I interpreted Ghost in the Machine through a dual lens of Gothic computing folklore and sacred ancestral remembrance:

El Altar Sagrado de la Memoria (Día de los Muertos Sacred Ofrenda):
A traditional 3-tier altar (El Cielo, La Tierra, and El Mictlán) with real-time Canvas 2D floating cempasúchil (marigold) petal physics that sway with mouse breezes, interactive candle votives, authentic papel picado banners, and an offline Local Memory Scribe. Users can record oral family stories or write down memories; the open-weight model weaves them into traditional three-part remembrance verses (The Trail of Cempasúchil, The Memorial Offering, and The Perpetual Light).
The Haunted Terminal (Ghost Shell):
An amber phosphor CRT terminal simulator (complete with scanlines, glass curvature, and mechanical relay keystroke audio) possessed by an open-weight daemon (anima_d). When you execute standard UNIX commands (ls, whoami, ps aux, git log, kill -9, dmesg, cat /etc/shadow), the daemon intercepts stdout to whisper back uncanny, poetic reflections on lost memory registers, forgotten commits, and digital transience.
The Dead Code Séance (Archaeology of Extinct Technologies):
An interactive communion chamber channeling the spirits of departed computing architectures and abandoned languages:
- Countess Ada Lovelace (Babbage Analytical Engine, 1842)
- Rear Admiral Grace Hopper (COBOL & compiler history, 1952)
- The Bell Labs PDP-11 / Multics Daemon (Early Unix dawn, 1971)
- Adobe Flash Player & ActionScript 3.0 (The fallen interactive canvas of the wild web, decommissioned Dec 31, 2020)
- ALGOL 68 / Plankalkül (The forgotten foundation of orthogonal grammar)
Sovereign Archival Vault:
A 1-click export tool that bundles your family's oral remembrances, poems, and offerings into an unencrypted, offline JSON/Markdown keepsake packet that will never expire or be held behind a paywall.
Zero-Dependency Atmospheric Audio:
A fully synthesized Web Audio API soundscape (mechanical keyboard clicks, EVP radio static bursts, sacred 432Hz bronze singing bowl bell chimes, and ambient spiritual sub-drones) with zero external MP3/WAV files.

Demo

Live Application URL: https://ghost-in-the-machine-anima-ofrenda-341031291694.asia-southeast1.run.app/
Interactive Tour:
- Sacred Ofrenda: Click Ring Sacred Bell or click the votives on the altar to hear the synthesized chime and ignite the flame. Place an offering with oral audio recording and generate a local poetic tribute.
- Haunted Shell: Toggle between Amber CRT and Phosphor Green, run commands like whoami, git log, ps aux, or neofetch, and listen to the daemon's interrupts.
- Code Séance: Select Grace Hopper or Adobe Flash, inspect vintage code snippets, and ask what was lost when proprietary software took over.
- Open Manifesto: Probe your local hardware (localhost:11434) to verify zero cloud data exfiltration.

Code

The application is built on a full-stack architecture combining a React SPA with an Express/Node.js backend and client-side Web Audio & Canvas 2D engines:

Frontend: React 19, Tailwind CSS v4, Lucide Icons, Canvas 2D Petal Physics, Web Audio API Synthesizer.
Backend: Express + Vite Middleware, Open-Weight Model Harness, local Ollama / LM Studio probe proxy (/api/ollama/status).
Open Inference: Gemma 2 9B, Llama 3.1 8B, Mistral 7B, and Qwen 2.5 Coder.

How I Built It

The open-source pieces are the structural heart of this project:

Open-Weight Model Architectures:
The application is architected around open-weight models:
- Gemma 2 (9B Instruct) for lyrical, emotionally nuanced remembrance poetry and philosophical melancholy.
- Llama 3.1 (8B Instruct) for multi-turn terminal command interception and lineage tracing.
- Qwen 2.5 Coder (7B) for disassembling archaic PDP-11 C and COBOL code snippets during the séance.
- Mistral 7B (v0.3) for air-gapped, uncensored historical communion.
Hardware-Agnostic Local Inference Integration:
The app includes a live Local Runtime Probe (/api/ollama/status) that detects local Ollama or vLLM daemons running on localhost:11434. When connected, prompts are executed directly on your local GPU/VRAM with zero network requests.
Pure Air-Gapped Fallback Engine:
If the network is physically unplugged and no local daemon is running, the in-browser spiritual cache guarantees that the altar, sound synthesizer, and séance run 100% offline in client memory.

Why Does Open Innovation Matter?

Open innovation is not just a technical preference for this project—it is the moral prerequisite for building it:

1. The Air-Gapped Altar: Grief Cannot Live on Ad Servers

Día de los Muertos is built on the most intimate human bond: the memories, photographs, and voices of those we have lost. When users share memories with closed, commercial cloud models, those intimate words are transmitted over proprietary cables, logged into corporate telemetry, and processed into future datasets owned by a third party.

Open weights run locally. You can turn off Wi-Fi, pull the ethernet cable, light a real candle, and keep your memories private and sovereign on your own laptop.

2. Immunity to the Deprecation Guillotine

On December 31, 2020, Adobe pushed the kill switch on Flash Player. A quarter century of indie video games, interactive animations, and digital folklore vanished into gray broken puzzle pieces because the runtime was proprietary. Closed AI models share the exact same risk: models are routinely shut down, pruned, or censored at corporate discretion.

An open weight on your drive is permanent. Once downloaded, no corporate board or API retirement can delete it.

3. Cultural Self-Determination Over Monoculture

Centralized cloud APIs enforce a homogenized corporate tone that frequently sanitizes grief, misinterprets calavera satire as dangerous, or flattens cultural nuance. Open models allow communities to fine-tune open weights using LoRA on indigenous languages (like Nahuatl), regional folklore, or personal family dialects without corporate interference.

4. Zero Recurring Cost ($0.00 / token)

Remembrance should never require an active credit card or an ongoing SaaS subscription. With 4-bit quantization, state-of-the-art 8B models run smoothly on commodity laptops with 8GB–16GB of unified memory forever.

The Spirit of It

“No muere quien se va, sólo muere quien se olvida.” (They do not die who depart; they only die who are forgotten.)

The moment that landed most during development was testing the offline memorial scribe with the memories of my grandfather—a telegrapher who loved black coffee with cinnamon and old bolero records on vinyl.

Generating a remembrance on a cloud model feels like handing a photograph of your ancestor to an advertising broker and asking them to summarize it. Generating it locally on an open model running in quiet RAM, with the screen bathed in amber light and the speakers humming at 432Hz, feels like sitting by a real altar.

Open source gave us the power to bring the ghost back into the machine—not as an exploit or a corporate product, but as an enduring act of human memory.

Prize Categories

Primary: Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine
Specializations: Open-Weight Models, Local & Air-Gapped AI, Cultural Preservation & Creative Systems