AREMES HQ, Brooklyn, May 25th 2026
Today I spent nearly an entire day inside Terminal on my macOS building an experimental semantic archive intelligence system around my lifelong WordPress media library. This was raw terminal-based systems building in collaboration with my friend Sir Claude Code, running locally through Node.js, Ollama, WordPress REST APIs, vector embeddings, semantic clustering systems, and custom archive intelligence tooling.
The entire process unfolded live through hundreds of terminal operations, syntax checks, vector validations, ingestion passes, embedding pipelines, cluster analysis runs, semantic nearest-neighbor generation, static export systems, and archive intelligence reports.
At multiple points the machine appeared less like a search engine and more like an archaeological system excavating hidden structures from twenty years of accumulated visual output. For the last few years I have been thinking deeply about a strange problem that I feel almost nobody talks about, ever.. What happens when a person has been publishing creative work to the internet continuously for over twenty years? I cant even imagine that this much time has even passed.. but it has indeed.
This was not casually posting, not optimizing for trends, not building for algorithms. Actually publishing. Consciously.
Thousands and thousands of artworks, drawings, animations, experiments, scans, paintings, GIFs, photographs, sculpture, prints, collage, prototypes, motion studies, AR/VR tests, 3D models, abstractions, video art, Internet Art, installations, tutorials and fragments of process spread across WordPress, GIPHY, cloud drives, external hard drives, old websites, Tumblr-era internet culture, and multiple generations of digital platforms.
At a certain point the archives become too large for chronology to mean anything. I’m a WordPress guy. I fell in love with it from the day that I learned about it in 2004. I watched from the sidelines for a year and half and then I jumped in, launching my first site in 2006. I don’t believe that WordPress media libraries back in 2026 were designed to function as intelligent cultural systems. They are essentially giant chronological storage buckets. The deeper the archive becomes, the more invisible the work becomes. Search breaks down. SEO becomes increasingly unreliable. Older work disappears beneath newer uploads. Valuable relationships between works are never surfaced.
An archive eventually becomes unreadable. This became daunting. Im a high volume production kind of artist. Im constantly making new things, everyday. I document those things, everyday. Im also Deaf and Hard of Hearing and I learn almost everything from visually reverse engineering things into some tangible example. But again, the archive became an abstraction, a real problem and I wanted to solve it.
This is not “AI art”, “AI content generation”, or another chatbot.
I wanted to know if an AI system could semantically understand a lifelong creative archive? One with just under 10K worth of artwork images, multidisciplinary images..
And more importantly, can it reorganize the archive into something discoverable again?
That became the foundation of what evolved into the AREMES Archive OS.
The Archive
The test archive was my own WordPress media library from ryanseslow.com
The domain and site has been active for well over seventeen years and currently contains approximately:
- 9,386 publicly accessible media records
- 20 years of accumulated visual output
- paintings
- drawings / illustration
- sculpture
- animated GIFs
- motion graphics / animation / video art
- photography
- 3D models (glb/usdz)
- PDFs / docs / written suchness
- visual fragments / Internet art
- experimental AI works
- spatial computing tests
- AR/VR prototypes
The important thing is that the archive was real. This was not a clean startup dataset. This was not a curated museum database.
This was not a demo collection. It was a living archive with all the messiness that real creative production accumulates over decades.. a total mess.
The Goal
The goal was to build a local semantic archive engine capable of:
- ingesting WordPress media libraries
- generating embeddings
- performing semantic search
- clustering related works
- identifying nearest neighbors
- surfacing hidden relationships
- generating archive intelligence reports
- eventually powering licensing, discovery, and curatorial systems
Importantly, I wanted the system to remain:
- read-only
- local-first
- resumable
- portable
- inexpensive
- API-driven
- WordPress-native
- deployable without complex infrastructure
No giant cloud stack. No venture-funded infrastructure (though that would be so nice!) No dependency-heavy AI startup architecture. Just intelligent archival systems built directly on top of existing cultural output.
The Tech Stack
The system was built primarily as a Node.js CLI application.
Core stack:
- Node.js
- vanilla JavaScript
- local JSON pipelines
- WordPress REST API
- Ollama
- nomic-embed-text embeddings
- cosine similarity vector search
- static HTML/CSS/JS export architecture
- Terminal / MacOS
- Claude Code
- Chat-gpt
The entire system intentionally avoided:
- databases
- vector databases
- cloud GPU infrastructure
- SaaS dependencies
- server-side runtime requirements
Everything operated through local flat-file architecture.
The archive lived primarily inside JSON artifacts:
- media_archive.json
- media_embedding_corpus.jsonl
- media_embeddings.jsonl
- clusters.json
- nearest_neighbors.json
- archive_intelligence.json
The entire system was effectively building a semantic operating layer over a WordPress archive.
The First Breakthrough: Semantic Search Actually Worked
The first major validation happened during vector testing. A semantic query was run against embedded works:
“dimensional graffiti sculpture entity”
The lexical search results were terrible. Only literal keyword matches appeared. But once vector similarity was enabled using real nomic embeddings through Ollama, the system began surfacing semantically related works that shared no direct keyword overlap.
It pulled:
- bronze/graffiti hybrid forms
- volumetric character sculptures
- 3D spatial abstractions
- hybrid graffiti entities
- sculptural motion studies
That was the moment the project became real. Excited! (I was already hours in!)
The archive was no longer searching by words. It was searching by meaning.
Embedding the Archive
The next stage involved embedding the archive itself.
The system successfully:
- paginated through 97 WordPress API pages
- ingested 9,386 media records
- regenerated archive corpus files
- preserved existing embeddings safely
- resumed embeddings incrementally
- validated semantic relationships
Initial semantic coverage:
- 500 embedded works
- 679 validated vectors across both ryanseslow + aremes
- 75 semantic clusters
- 3 large semantic “worlds”
- multiple emergent series and collections
The system identified:
- recurring visual motifs
- medium transitions
- temporal shifts
- outlier works
- semantic neighborhoods
- 2D → 3D transformation relationships
One particularly fascinating discovery was how often photography re-emerged across decades despite enormous stylistic variation.
The archive was beginning to reveal patterns that were difficult to recognize chronologically.
The Clustering Experiments
One of the strongest moments of the process was the semantic clustering layer. Instead of manually tagging works, the system grouped works through vector proximity and centroid similarity.
Clusters began emerging naturally:
- sculptural portrait systems
- 3D spatial hybrids
- animation worlds
- museum/digital abstractions
- collage systems
- glitch structures
- graffiti-derived volumetric forms
Some clusters were extremely coherent. Others collapsed into noise. That became one of the most important realizations of the entire experiment:
Semantic similarity does not automatically equal aesthetic coherence..
AI can recognize relationships. But curation still matters.
The Archive Intelligence Layer
The archive-intelligence mode became one of the most ambitious parts of the build.
The system joined:
- archive metadata
- embeddings
- cluster relationships
- nearest-neighbor systems
- temporal analysis
- semantic series
- cross-medium relationships
It generated:
- semantic collections
- inferred exhibition titles
- neighboring works
- outlier detection
- motif analysis
- “world” structures
- licensing potentials
- spatial potentials
At this stage the system was no longer simply indexing media. It was beginning to behave more like a curatorial intelligence layer.
The Most Important Realization
After several hours of successful backend engineering, an important realization appeared:
A CLI has no buyer. (Funny.. and not funny!)
That sentence completely changed the direction of the project. (I had been slurped in, once again, but I love that!)
The engine worked. The semantic systems worked. The archive intelligence worked. But nobody could see it. Everything still lived in terminal windows and JSON files. The project had become an extremely sophisticated invisible machine.
That forced a much bigger question:
What is the actual public-facing surface?
The Export-Site Experiment
The next phase attempted to solve this problem. A static semantic archive site was generated directly from the JSON outputs.
The idea was powerful:
- semantic discovery
- related works
- cluster navigation
- curated series
- licensing CTAs
- semantic search
- archive worlds
The system generated:
- index.html
- style.css
- app.js
No backend. No runtime AI. No database. No server dependency (perhaps I try to deploy on wordPress Sandbox?) Just a static semantic archive generated from the intelligence layer. Conceptually, this was exactly the correct direction. Visually, however, the system immediately exposed another difficult truth.
The Failure That Mattered Most
The semantic engine worked. The visual orchestration did not!
The archive surface became visually unstable:
- mixed image ratios
- broken previews
- inconsistent media sizes
- GIF chaos
- missing thumbnails
- 3D objects
- PDFs
- wildly different eras colliding together
The result was technically impressive but aesthetically uneven. And honestly, that failure may have been the most important discovery of the entire day. Because it clarified something critical:
AI-generated archive systems still require human taste. Semantic relationships are not enough.
Museum-grade experiences require:
- pacing
- hierarchy
- rhythm
- restraint
- spatial composition
- curatorial intelligence
- emotional sequencing
This was the exact point where the project shifted from backend engineering to art direction..
The Real Opportunity
The deeper realization is that the semantic engine itself is not the product. The archive IS the product.
The engine becomes:
- the curator
- the navigator
- the merchandiser
- the discovery layer
- the licensing assistant
- the relationship engine
That distinction changes everything.
Because suddenly:
- older works become discoverable again
- semantic relationships become visible
- licensing becomes easier
- collections emerge automatically
- AI agents can traverse the archive meaningfully
- archives stop behaving like dead storage systems
This is especially important for artists, museums, photographers, designers, institutions, universities, and cultural archives with decades of accumulated digital material.
Why This Matters Beyond My Own Archive
Most WordPress media libraries are dormant semantic archives. Millions of people have already unknowingly built enormous cultural datasets. The problem is, those archives are largely unreadable.
This experiment suggests another future:
- semantic museum systems
- agent-readable archives
- intelligent licensing discovery
- AI-assisted curatorial navigation
- AR/VR semantic galleries
- spatial archive interfaces
- archive intelligence layers on top of existing cultural systems
The important thing is that none of this required rebuilding the internet.
The entire system operated on top of:
- WordPress
- JSON
- local embeddings
- static exports
- open APIs
The architecture remained surprisingly lightweight.
What Happens Next
At this point the project has proven:
- semantic ingestion works
- embeddings work
- clustering works
- archive intelligence works
- export systems work
What remains unresolved is -> visual orchestration..
That is now the real frontier. Not “more AI.” Not larger models. Not more embeddings.
The challenge now is: how to transform semantic intelligence into elegant cultural interfaces. Yes, aesthetics, we like pretty things to look at..
That is a design problem as much as a technical one. Im up for it!
Final Thoughts
This entire experiment started with a simple question:
Can an AI system understand a lifelong archive?
The answer appears to be: yes, partially. But another question emerged underneath it: Can intelligence alone create meaning?
The answer to that is much more complicated…
Semantic systems can identify relationships. They can surface hidden structures. They can organize massive archives. They can discover patterns humans overlook. But they still cannot replace curatorial sensitivity, restraint, pacing, and aesthetic judgment.. right? Yet? Hmm..
The machine can understand proximity.. The human still understands significance..
And maybe that balance is still the actual future, I don’t know, but Im excited to find out, and continue to tinker. I don’t want AI replacing archives, but I do want AI making archives visible again.
