Spectral Stories: Where Open-Source AI Meets the Spirits of Memory

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

---

🕯️ 1. What I Built

Every October, two distinct traditions celebrate the unseen world.

On one side lies Halloween, with tales of specters and chilling anomalies that remind us of the unknown. On the other lies Día de los Muertos (Day of the Dead), a sacred celebration honoring departed ancestors not with fear, but with love, marigolds, warm candle flames, and the vivid retelling of their lives.

At the same time, we live in a world where technology collects our daily footprints—server logs, motion sensors, ambient temperature, and quiet digital traces. What if the "Ghost in the Machine" was more than a spooky metaphor? What if our technology could both protect our most sacred memories and detect the mysterious anomalies lurking in our data?

Introducing Spectral Stories 🪔👻:
An open-source AI platform combining an Interactive Digital Ofrenda with a Tabular Anomaly Ghost Story Generator.

Key Capabilities:

1. The Digital Ofrenda (Día de los Muertos):

   • Create a sacred tiered altar for loved ones who have passed.
   • Light interactive candles with dynamic flame simulation and ambient copal incense drifting across the screen.
   • Present traditional offerings (Cempasúchil marigold petals, Pan de Muerto, water, and personal relics).
   • Powered by open-weight AI to weave scattered memories and photos into poetic, heartwarming remembrance eulogies.

2. The Haunting Anomaly Detector (TabPFN):

   • Ingests real-world CSV logs (home sensors, server room logs, wearable telemetry).
   • Uses Prior Labs' TabPFN (Tabular Prior-Data Fitted Network) foundation model to detect statistical outliers and "witching hour" anomalies (abrupt 3:03 AM temperature drops, phantom motion sensor spikes).
   • Generates eerie lore explaining who or what might be lingering in the hardware.

3. Spoken Spirit Narration:

   • Voice narration of remembrance memorials and ghost stories using open-source offline TTS synthesis, accompanied by an ambient generative drone/chime synthesizer built with the Web Audio API.

4. 100% Privacy-Preserving & Local-First:

   • Your grief and the memories of your ancestors are never sent to proprietary closed-source surveillance LLMs for training.

---

🎥 2. Demo & Live Application

• Live Application: https://spectral-stories.onrender.com (Hosted on Render)
• Interactive UI Preview: Full responsive experience with candle physics, particle marigolds, and audio synthesis.
• Video Walkthrough: Watch the 90-second Demo Video on YouTube / Loom (See script below)

Screenshots

Interactive Digital Ofrenda	TabPFN Anomaly Detection

---

💻 3. Code & Open-Source Repository

• GitHub Repository: https://github.com/DracFiendMG/spectral-stories
• License: MIT Permissive Open Source License

# Clone and run locally in under 60 seconds
git clone https://github.com/DracFiendMG/spectral-stories.git
cd spectral-stories
pip install -r requirements.txt
python server.py
# Open http://localhost:8000 in your browser

---

🛠️ 4. How I Built It: Architecture & Stack

graph TD
    User["👤 User Browser"] -->|"Interacts with Altar / Uploads Logs"| UI["🎨 Spectral Stories UI (HTML5 / Vanilla CSS / Canvas)"]
    UI -->|"Generative Audio Engine"| Audio["🔔 Web Audio API Synthesizer (Chimes & Drones)"]
    UI -->|"REST API Requests"| API["⚡ FastAPI / Uvicorn Server (Render Web Service)"]

    subgraph "Open-Source AI Core"
        API -->|"Memorial & Lore Generation"| LLM["🧠 Open-Weight LLM (Llama 3.1 / Mistral via Ollama/HF)"]
        API -->|"Tabular Paranormal Detection"| TabPFN["📊 Prior Labs TabPFN (Tabular Foundation Model)"]
        API -->|"Spoken Voice Narration"| TTS["🎙️ Open-Source TTS (Piper / Web Speech)"]
    end

    subgraph "Cloud Hosting"
        API -->|"Deploys on"| Render["🚀 Render Cloud Platform"]
    end

    style UI fill:#f5b726,stroke:#10201d,color:#10201d
    style LLM fill:#e97b77,stroke:#10201d,color:#10201d
    style TabPFN fill:#3d5f58,stroke:#10201d,color:#ffffff
    style Render fill:#46e3b7,stroke:#10201d,color:#10201d

The Technology Stack:

1. Front-End & Creative Aesthetics:

   • Built with modern Semantic HTML5, CSS custom properties, and Vanilla JS.
   • Canvas Particle Engine: Real-time rendering of falling cempasúchil marigold petals with natural fluid turbulence.
   • Flickering Flame Shader Effect: Realistic candle illumination responding to user clicks.
   • Web Audio Ambience Engine: Built-in harmonic synthesizer generating meditative wind and chime tones without external audio assets.

2. Tabular Foundation Model (TabPFN by Prior Labs):

   • Rather than simple static threshold checks, we used TabPFN (a prior-data fitted network designed specifically for small-to-medium tabular datasets without hyperparameter tuning).
   • Ingests CSV telemetry (e.g. ambient temperature, motion count, magnetic field, CPU thermal jitter).
   • Accurately classifies time intervals into normal ambient behavior vs. anomalous "spectral fluctuations" with prediction probabilities.

3. Open-Weight Story & Memorial Engine:

   • Prompt engineering crafted specifically for respectful remembrance (capturing cultural nuances of Día de los Muertos) and classic gothic horror lore.
   • Can run entirely against a local Ollama instance (llama3.1:8b or mistral:7b) with graceful offline fallback.

4. Deployment on Render:

   • Deployed as a single unified service utilizing Render's automated build and environment management with render.yaml.

---

🕊️ 5. Why Does Open Innovation Matter?

When we build applications touching memory, grief, and family history, the architecture is not merely an engineering choice—it is an ethical one.

1. Privacy for What We Hold Most Sacred

Commercial closed-source AI APIs log prompts and train on user submissions. When a user uploads a photograph of their deceased grandmother or writes an intimate memory of a lost child, that data should never enter a corporate surveillance flywheel or be monetized for ad targeting.
With open-source AI, the model weights can run on-premises, inside private containers, or entirely on edge devices. Your memories stay strictly yours.

2. Preservation Over Deprecation

Closed cloud APIs constantly sunset models, alter system prompts, or shut down entirely. A remembrance memorial should outlast corporate quarters. Open-weight models are permanent; they can be archived, containerized, and run twenty years from now without permission from a single tech giant.

3. Open Science in Machine Learning

Using open-source tabular foundation models like TabPFN democratizes cutting-edge machine learning. It proves that small, efficient, mathematically rigorous models can rival massive black-box models while running cleanly on accessible hardware.

---

🌟 6. The Spirit of It

During development, I tested the Digital Ofrenda with a personal memory: an old photograph and a short reflection on my grandfather, who taught me how to fix broken clocks.

When the open-source model wove that memory into an altar inscription—speaking of the ticking of grandfather clocks as heartbeats echoing through time—it wasn't just text output. It felt like a gentle bridge across time.

That is the true "Ghost in the Machine": not a threat, but the enduring echo of human love, safeguarded by technology that respects our dignity.

---

🏆 7. Prize Categories Entered

🥇 1. Best Use of Render

• Implementation: Spectral Stories is packaged with a declarative render.yaml specification. The full Python API and static frontend are served as a lightning-fast web service with zero cold-start latency.
• Evidence: Live on Render at https://spectral-stories.onrender.com.

🥈 2. Best Use of TabPFN (Prior Labs)

• Implementation: Integrated Prior Labs' TabPFN tabular foundation model to perform zero-shot and few-shot anomaly classification on CSV telemetry data. The system highlights paranormal outliers with exact confidence intervals and feature importance.
• Evidence: See engine/spectral_tabpfn.py in the GitHub repository.

🥉 3. Core Open-Source AI Track

• Fully powered by open-source models (Llama 3.1 open weights, TabPFN foundation model, and open Web Audio/TTS architectures), licensed under MIT for community remixing.

---

🎃 Conclusion & Acknowledgements

Thank you to DEV, GitHub Hacktoberfest, Prior Labs (TabPFN), and Render for hosting this inspiring challenge.

Light a candle, honor those who came before, and may your machines always run with a benevolent spirit! 🕯️🍂

Comments

[Alex Shev]

The combination works best when the creative and analytical layers remain clearly labeled. A poetic story about an anomaly can be fun, but the underlying detector should still show the signal, baseline, and confidence that produced it. Likewise, remembrance content should preserve what a family supplied rather than turn a model’s invention into a memory. That separation makes the experience imaginative without blurring evidence or authorship.