Some failures disappear from memory the moment the system starts working again. The equipment gets repaired, the alerts stop, and the incident becomes another entry in an archive.
But what if the most important part of that incident was never the failed component? What if it was the chain of dependencies nobody fully understood?
That question inspired me to build GHOSTTRACE, an open-weight AI investigation laboratory designed to recover the hidden stories behind infrastructure failures.
What I Built
GHOSTTRACE transforms written incident reports into interactive investigations.
Instead of presenting users with another AI chatbot, I wanted to create an environment where evidence could be discovered, interpreted, questioned and preserved.
The application lets users explore incident archives, identify relevant evidence, reconstruct potential system dependencies and examine how disruptions might have spread.
I also introduced Memory Vigil, a remembrance experience inspired by Día de los Muertos. Visitors can light symbolic lanterns for documented traces, preserving the lessons that might otherwise disappear with time.
The included fictional incidents provide a safe way to explore the experience without presenting invented events as historical facts.
Demo
Live application: https://ghosttrace-six.vercel.app/
To experience GHOSTTRACE, open an incident archive, run the Evidence Lens, reconstruct the report with Claude, inspect the resulting network and visit the Memory Vigil.
Code
GitHub: https://github.com/paulos-ui/ghosttrace
The repository contains the application source, evidence-processing logic, reconstruction pipeline, graph engine and documentation.
How I Built It
GHOSTTRACE combines two different approaches to AI.
The first is an open-weight evidence model. Using Transformers.js and the Apache-2.0 licensed MiniLM model, the application generates embeddings directly in the browser.
These embeddings help identify relevant passages from the original report. Users can inspect the retrieved evidence and its relevance scores before any further reconstruction takes place.
The second component uses the Claude API for structured interpretation. Claude helps extract systems, possible dependencies, event sequences and unresolved questions.
Claude is proprietary, and I intentionally keep that distinction clear. The open-weight evidence retrieval works independently of Claude, while the hosted model provides optional structured analysis.
The reconstructed results can be explored through an interactive network. A deterministic graph engine supports hypothetical containment experiments, while evidence quotation checks help distinguish text found in the report from uncertain interpretations.
The interface is built with Next.js, React and TypeScript and deployed on Vercel.
Why Open Innovation Matters
For me, the most important decision was making evidence retrieval open and inspectable.
A closed AI service can generate a convincing explanation, but that explanation becomes much more useful when users can see which passages informed the investigation.
With MiniLM running in the browser, evidence retrieval does not require a proprietary inference API key. After the model is downloaded, the embedding calculations happen locally.
That also means developers can inspect the implementation, experiment with alternative models and change how evidence is retrieved.
The approach has limitations. Initial model downloads require internet access, and choosing Claude reconstruction sends report content to a hosted API. GHOSTTRACE makes those distinctions explicit rather than claiming that the entire application runs offline.
Open innovation matters here because the evidence-discovery layer can remain independently useful, transparent and replaceable.
The Spirit of It
I interpreted Ghost in the Machine in two ways.
For Halloween, GHOSTTRACE presents forgotten incidents through the atmosphere of a haunted technical archive. A failed system becomes a mystery, and the remaining evidence becomes the trail that leads us through it.
For Día de los Muertos, I wanted to explore remembrance more thoughtfully.
The Memory Vigil gives documented traces a place to remain visible through symbolic lanterns and evidence excerpts.
It does not pretend to resurrect people, recreate memories or invent historical testimony.
It preserves what the available evidence can actually tell us.
What I Learned
The hardest part was not designing a spooky interface. It was deciding how to prevent an impressive-looking reconstruction from becoming more convincing than its evidence.
Two events occurring close together do not automatically prove one caused the other.
That distinction shaped the application. Evidence ranking, AI interpretation, quotation verification and deterministic simulations remain separate parts of the experience.
GHOSTTRACE is not a certified disaster-prediction system. It is a working exploration tool built to make forgotten dependencies easier to investigate and understand.
The project reminded me that good AI development is not only about generating answers. Sometimes it is about making uncertainty visible and preserving the evidence worth remembering.
GHOSTTRACE. Every system leaves a ghost.
Top comments (0)