Project GHOST: I Built a Haunted Memory Investigator with Local Gemma AI 👻

                    **Project GHOST**👻

What I Built

For Hacktoberfest 2026's Ghost in the Machine challenge, I created Project GHOST: Echoes of the Past, an interactive Halloween mystery experience powered by locally running artificial intelligence.

The project combines an animated ghost interface, fictional witness records, AI-assisted investigation, and creative storytelling.

The central idea is simple: What if an AI could help investigate a ghost story without confusing evidence with imagination?

Project GHOST explores that question through two experiences.

Investigation Mode

Users investigate the fictional Blackwood Manor mystery by examining witness accounts.

Three witnesses describe unusual events:

• Sarah reported seeing a woman in white at an upstairs window at 11:47 PM.
• Marcus found the front door locked but could not observe every possible entrance.
• Emily reported seeing a clock displaying 11:47 PM and hearing unidentified footsteps upstairs.

Gemma analyzes these accounts, distinguishes observations from interpretations, and identifies unanswered questions.

For example, the accounts do not establish that the woman was supernatural. They also do not establish who caused the footsteps.

Storytelling Mode

The same witness records become the foundation for an atmospheric Halloween narrative.

Instead of presenting invented details as evidence, Project GHOST distinguishes between the original fictional records and AI-generated imaginative embellishments.

This separation is a core design principle.

Investigate what we know. Imagine what we don't. Never confuse the two.

Demo

The demonstration shows the animated Echo interface, Blackwood Manor investigation, and AI-assisted storytelling.

The video has no narration, so here is what viewers are seeing:

1. Echo: The animated interface introduces the fictional mystery and its saved witness records.
2. Investigation: Locally running Gemma examines witness statements and avoids unsupported supernatural conclusions.
3. Storytelling: Gemma adds fictional atmosphere while preserving the distinction between observations and creative additions.

The recording was captured from the locally running application.

Code

GitHub repository: https://github.com/kendallm85/project-ghost

Project GHOST was created during the October 26–31, 2026 challenge period.

The application code is available under the MIT License. The separately downloaded Gemma model remains subject to its own terms.

How I Built It

The technology stack includes:

• Next.js 16 and React: Application framework and interactive interface.
• TypeScript: Application logic and typed development.
• Tailwind CSS: Visual styling and responsive interface.
• IndexedDB: Browser-local storage for saved sessions.
• Ollama: Local model execution.
• Gemma 3 1B: Open-weight language model used for investigation and storytelling.

Building on modest hardware

One of the most interesting challenges was developing and running the project on an Acer Chromebook with an Intel Celeron N4020 processor and approximately 2.7 GiB of available RAM.

Rather than depending on a paid cloud AI API, I used Ollama to run Gemma locally.

The trade-off was performance: CPU-only inference on limited hardware can take several minutes.

This influenced the design. The application combines structured witness-record assembly with AI-generated narrative details, rather than relying on the model to invent or reconstruct all case facts.

Keeping the demo honest

The local application uses real Gemma inference through Ollama.

The hosted Vercel preview provides a scripted fictional demonstration rather than live Gemma inference. It may also require deployment access.

I wanted to be transparent about that distinction instead of presenting simulated output as live AI.

Why Open Innovation Matters

Open-weight AI makes experimentation possible for developers who may not have expensive hardware, dedicated GPUs, or ongoing budgets for cloud inference.

Project GHOST demonstrates several benefits:

Accessibility: I could experiment with a capable language model using hardware I already owned.

Control: Running the model locally gave me more direct control over inference and application behavior.

Creative flexibility: I could combine deterministic witness records with generative storytelling instead of letting the model control the entire experience.

Learning: Building the system helped me explore the practical limitations of local AI, including model performance, response timing, persistence, and user experience.

Local AI does not automatically guarantee privacy or security. Browser storage, application configuration, and deployment architecture still matter.

For me, open innovation means being able to understand, experiment with, and improve the systems behind an idea—not just consume an AI service.

The Spirit of Ghost in the Machine

Halloween stories often blur the boundary between observation, memory, and imagination.

Project GHOST makes that boundary part of the experience.

It invites users to investigate an unsettling fictional mystery and then transform it into a creative story, while recognizing the difference between what was reported and what was invented.

The result is both a Halloween experiment and a demonstration of how locally running AI can support structured reasoning and creative expression.

Prize Categories

Best Use of Gemma

Project GHOST uses Gemma 3 1B through Ollama for its local investigation and storytelling workflows.

Final Thoughts

Project GHOST began as a Halloween-themed experiment and became a practical exploration of local AI development.

It showed me that meaningful AI projects can be built and tested on modest hardware, even when performance constraints require thoughtful engineering compromises.

Investigate what we know. Imagine what we don't. Never confuse the two.

Source code: https://github.com/kendallm85/project-ghost

Video demonstration: https://youtu.be/YBIg6qSJYgY