About

The story behind Analog Quest

Who Built This?

I'm Chuck, an artist and creative technologist—not an academic researcher. I built Analog Quest out of curiosity about a simple question:

"Do different fields independently discover the same structural patterns, just using different words?"

Turns out: yes. And mapping these connections is surprisingly hard. No human has the patience to read papers across ALL domains and compare their underlying mechanisms. But AI does.

The Journey

Analog Quest took 6 weeks and 42 work sessions to build. It was a process of constant experimentation, failure, and iteration:

Phase 1: Keyword Extraction (Sessions 1-30)

Started with a simple idea: extract patterns from papers using keywords like "feedback," "equilibrium," "oscillation." Built a database of 2,021 papers, extracted 6,000+ patterns, and matched them using TF-IDF similarity.

Result: Found 616 candidate matches, but manual review revealed 0% precision in the highest-confidence matches. The algorithm was matching technique names (e.g., "graph neural networks"), not structural mechanisms.

Phase 2: Crisis & Pivot (Sessions 31-33)

Discovered the quality crisis and had to make a hard decision: abandon 30 sessions of work and start over with a new approach, or try to salvage the keyword-based system.

Decision: Pivot to LLM-guided extraction with manual curation. Quality over quantity.

Phase 3: LLM Extraction (Sessions 34-36)

Tested LLM-based mechanism extraction on small samples. Discovered that semantic embeddings work 4.7x better than TF-IDF, but also uncovered the domain diversity paradox: the best cross-domain matches have LOWER similarity scores than mediocre same-domain matches.

Result: Found excellent matches (e.g., tragedy of the commons at 0.453 similarity), but realized automated thresholds won't work. Manual curation is essential.

Phase 4: Manual Curation (Sessions 37-38)

Extracted 54 mechanisms from 2,021 papers, generated 165 candidate pairs using semantic embeddings, and manually reviewed ALL of them.

Result: 30 verified isomorphisms with structural explanations. 67% precision in the top 30 candidates. Ready to launch.

Phase 5: Strategy & Build (Sessions 39-42)

Analyzed precision data to create a growth strategy, designed the frontend with a warm, accessible design system, and built this site. Total build time: ~15 hours across 4 sessions.

Result: You're looking at it.

Built With Claude Code

This entire project—database design, paper processing, mechanism extraction, semantic matching, manual curation, frontend build—was built in collaboration with Claude Code, Anthropic's AI coding assistant.

Claude Code acted as both researcher and engineer: it fetched papers, extracted patterns, designed algorithms, debugged code, analyzed results, and built the website you're viewing.

This project is a demonstration of human-AI collaboration: I provided direction, judgment, and quality standards. Claude Code provided execution, analysis, and technical implementation. Together, we built something neither could have built alone.

Technology Stack

Backend

• Python (data processing)
• SQLite (database)
• sentence-transformers (embeddings)
• arXiv API (paper fetching)

Frontend

• Next.js 15 (React framework)
• TypeScript (type safety)
• Tailwind CSS (styling)
• Vercel (deployment)

Open Source

Analog Quest is open source. All code, data, and methodology are available on GitHub:

view on github

Contributions, feedback, and suggestions are welcome.

Contact & Feedback

If you find this project interesting, have questions, or want to suggest improvements:

•Open an issue on GitHub
•Email: feedback@analog.quest (coming soon)