WPT–06 / Operating system

AI doesn't do my work.
It makes my work survive scrutiny.

I run every project through the same loop: AI as a thinking partner going in, and as a hostile reviewer coming out. Click any waypoint on the loop — each stage carries receipts from systems I've actually shipped.

the loop BUILD → BREAK → PROVE

WPT–07 / Architecture receipts

The loop, in production

Three shipped systems where the review-refine-verify stages are built into the architecture itself — watch the data move.

TravelGraph-AI — triple validation

Nothing the LLM writes touches the database until it survives three gates. Winner, KGC 2026 poster session.

3D exploded system · drag to rotate · scroll to zoom · click to assemble/explode
User question natural language Llama 3 drafts Cypher Gate 1 · schema does the graph agree? Gate 2 · syntax valid Cypher only Gate 3 · semantics answers the question? Neo4j KG 57,877 nodes Answer grounded invalid → blocked & regenerated, never executed

H.O.T Jarvis — the skill forge

The assistant writes its own new skill, writes the test that could kill it, and only keeps what passes.

ShellTauri v2 · React HUD · command palette
→
Model routerOllama local → Groq → OpenRouter free tiers
→
Skill forgeSandboxed code + the self-written test that gates it
→
MemorySQLite · versioned migrations · export / wipe
"Build me a skill" user intent Code + test written together Sandbox run prove it works Pass? tests are the gate Skill library versioned Flagged refuses to run fail → refine and re-forge

QuizLab — guarded RAG

Retrieval keeps it grounded in the syllabus; a prompt-security layer assumes the input is hostile.

IngestStreamlit · slides / URLs in
→
PromptShieldInjection defense — input assumed hostile
→
RAG engineLangChain · OpenAI embeddings · ChromaDB
→
GradingSemantic rubric — meaning, not regex
Slides / URLs course material Chunk + embed ChromaDB PromptShield injection defense Quiz LLM grounded generation Quiz syllabus-aligned Semantic grade meaning, not regex hostile input stops at the shield — never reaches the model

WPT–08 / Field rules

Three rules I don't break

Trust is earned per answer.

An AI response is a draft until it's verified — against the schema, the test suite, or a human rubric. Every system I ship has a gate the model must pass.

Real numbers only.

H.O.T Earth's README says it best: "real physics, real data, no fake numbers." If a metric can't be traced to a run, a log, or a repo — it doesn't ship.

The first answer is never the answer.

Re-prompt, re-frame, attack your own output. The loop exists because iteration — not generation — is where the quality comes from.