Return of the Coder
They said AI would replace coders. I used AI to build seven production systems that don't need AI to run. Here are the numbers.
All seven systems are active works in progress. Every number below is real and verified; each system is still being refined and extended.
91.4% UAS on Universal Dependencies — statistically tied with Stanford's Stanza neural parser (91.7%) on 6,125 sentences. 10 MB. Parses 6,125 sentences in 25 seconds, about 29× faster than Stanza. Zero neural inference at runtime. Runs in your browser.
Open the demo →A rule-based composer in the style of Mozart, C.P.E. Bach, Kirnberger. 92% statistical fidelity to Mozart originals across 20+ objective metrics. Full browser-side export: sheet music, PDF, MIDI, WAV. 264 tests passing. 0 LLM at runtime.
Play the composer →MYCIN-style reasoning engine with forward chaining, certainty factors, and a full justification graph. Every recommendation traces to specific rules. Exactly four LLM touchpoints, each a typed Semantic Function. Everything else deterministic. Bring your own API key.
Open the system →Twenty themes, forty-plus section components. Every build produces an HTML site, PDF, DOCX, plain-text resume, VCF card, OG image, sitemap. Deterministic. One-second builds. No framework runtime, no server.
Open the gallery →Thirty themes, twenty-four section components, thirteen gallery layouts. Outputs: landing page, per-work pages, printable catalog, PDF portfolio, ZIP archive, sitemap, RSS. One-second builds. No framework, no database, no server.
Open the gallery →A deterministic composer that turns a short feature description into a playable single-file HTML5 idle game. 600+ expert-system rules across 21 decision domains, 80+ presets spanning cozy management to eldritch deep to cyberpunk. Auto-tuning via Monte Carlo simulation. Every choice explainable and traceable to a specific rule firing.
Open the gallery →A DM's co-pilot that turns a short brief into a complete session-by-session campaign binder. 205 rules with MYCIN certainty factors, 10 orchestrated factories (campaign, NPCs, relationships, plot facts, contingencies, handouts, encounters, maps, coherence audit), 84 SRD 5.1 stat blocks. Seeded, deterministic, fully offline.
Open the campaigns →Each system started with a gold file — expected correct outputs written by hand. I asked an AI coding assistant to write a rule-based implementation that passed every case. Then iterated: rule, test, verify no regression on held-out data, commit. When it worked, I shipped a program. The AI is gone. The code remains.
This is not a trick. It is the method the pre-neural AI researchers spent half a century perfecting. They were right about what rule-based systems could do. They were wrong about one thing: who could write the rules. The knowledge-acquisition bottleneck that collapsed their field is the one thing AI is now genuinely good at dissolving.
The AI companies spent hundreds of billions of dollars building the most powerful knowledge-recombination engine in history. They told coders their careers were over. The best use of that engine turns out to be building systems that don't need it. Every gold file extracted is a domain the AI companies can never charge for again. Every tested library component is a piece of intelligence that runs on ordinary hardware, offline, forever.
The career that was supposed to die is the career best positioned to reshape the world — because it is the only career with the skill set to do the extracting.
I'm turning this essay into a book.
I'm publishing the library as I go.
Bookmark this page. New chapters and systems land here.