Skip to content
Contents

1. Revenue systems

2. Builds

3. The operator

4. Writing

5. Journal

6. Resume

7. Appendix

378 wordsby CONNOR J. LAUGHLIN

The wiki humans and agents share.

A knowledge base compiled once from the company's sources, with provenance, an owner, and an approval tier on every page.

Outcome

Built a governed LLM wiki on Karpathy's compile-once pattern. Humans and AI agents read the same pages for RFP response, content, internal FAQ, and agent context.

Compile-once pathThe governed LLM wiki's pipeline: the source corpus is extracted, compiled into interlinked machine-readable articles, and verified before a build ships. The two converging callouts are the point of the system: human readers and AI agents read the same governed pages.SOURCE CORPUSEXTRACTCOMPILEVERIFYGOVERNED WIKIHUMAN READERSAGENT READERS
FIG_024 [ Compile-once path ]Compiled once, verified, then read by humans and agents from the same pages.

Standard RAG re-reads the raw pile on every question and trusts whatever it grabs. I built the other thing, on Karpathy's compile-once pattern: the company's source records compiled once into interlinked, machine-readable articles with an index, then maintained by governed refresh builds.

People browse it as documents. Agents read the same files for context, under written instructions that tell them to start there. One set of pages, two kinds of reader.

The case

  1. 01The problem

    Every function kept re-deriving the same answers from scattered sources: proposal responses, content, internal FAQ, and now agents that need context. The knowledge existed, but it lived in thousands of records nobody could safely hand to an AI.

  2. 02What I built

    A governed LLM wiki: an extraction pass over the source corpus, a compiler that builds interlinked articles across the company's knowledge domains, and a verifier that stops any build with broken links or bad source IDs. Every article carries provenance, an owner, a verifier, and an approval tier from external-safe to internal-only, with a canonical-facts registry underneath.

  3. 03What changed

    Humans and AI agents now start from the same governed pages. The wiki feeds RFP response, content, internal FAQ, and agent context across the business.

  4. 04Why it mattered

    Pointing an agent at raw sources means re-deriving truth on every query. Compiling once means the organization argues about a page, approves it, and everything downstream inherits the decision.

  5. 05What it proves

    Before I point an agent at work, I build the knowledge it needs: governed, indexed, machine-readable, and shared with the humans doing the same job.

Proof

Chapter figures

interlinked articles
157. interlinked articles. machine-readable pages across 18 domain folders, compiled in one pass and maintained by refresh builds.
source records compiled
3,300+. source records compiled. the corpus the wiki is built from, extracted and chunked for retrieval.
canonical facts in the registry
1,772. canonical facts in the registry. mined from the most-reused proposal answers, each with a stable ID and a status tag.
broken links at build
0. broken links at build. build verification holds a release until wikilinks and source IDs check clean.

Systems built

  1. Corpus extraction over the company's source records
  2. Wiki compiler producing interlinked, machine-readable articles with per-article frontmatter
  3. Build verifier and a governed refresh flow that archives the prior build and preserves sign-offs
  4. Governance layer: canonical-facts registry, claims and proof registry, approval gates, source register
  5. Fact-vault intake that consolidates reused proposal answers into canonical facts
  6. Agent read instructions that point AI consumers at the wiki first

Chapter details

Scope
  • Corpus extraction
  • canonical facts
  • compiled wiki
  • verification
  • governed refresh.
Stack
  • Python
  • Markdown + YAML frontmatter
  • Wikilinks
  • Agent context files
Governance
  • Provenance and owners
  • approval tiers
  • canonical-facts registry
  • verification ledger

Governance notes

  • Architecture and adoption only: wiki contents stay internal
  • No usage counts and no answered-proposal claims; the build's own self-assessment records the external-approval ledger as unexercised
  • Every article names its owner, verifier, sources, and approval tier

In the interview

I compile the company's knowledge once, govern it, and point people and agents at the same pages.
Email Connor

Open to VP of Marketing & GTM, Head of GTM, VP Revenue Operations, and GTM engineering roles at AI-native B2B SaaS and PE-backed growth companies that build the tools of change. Chicago, hybrid or remote.