After 13 years as Director of Application Development at AANP, I resigned in October 2025. A month later, they brought me back as a contractor. The ask was straightforward: transfer everything in my head into something the team could actually use after I was gone.
That's a deceptively hard problem. Thirteen years of context - why systems were built a certain way, which vendor quirks to watch for, where the bodies are buried in the data model - doesn't fit neatly into a wiki. Especially when the organization is mid-migration from a legacy AMS to Salesforce and half the infrastructure is still in flight.
I authored over 160 technical documents (Salesforce architecture, payment processing, integrations, infrastructure, vendor support references, architecture diagrams, and third-party systems) all built on Zensical (a fork of MkDocs) with a custom style guide, YAML frontmatter standards, and Mermaid diagrams. But the part I'm most proud of is what happened when I realized traditional documentation wasn't going to be enough. I built an AI-powered knowledge base using Cloudflare AutoRAG with six separate indexed knowledge bases covering metadata, code, data migration scripts, business requirements, working notes, and processed video transcripts. Staff could ask questions in plain language and get cited answers, with source hierarchy weighting so they'd know whether they were reading authoritative documentation or a demo recording.
The video piece was its own project. When I resigned, a team member gifted me a Jetson Orin Nano Super (a tiny AI computer) and joked that I should use it to build "JeffGPT." So I did, more or less. I ran 230+ training recordings through a local Whisper transcription pipeline on that little machine, then used the Anthropic API to convert raw transcripts into structured documents with topic extraction, confidence tiers, and YAML metadata. Those got indexed separately so the AI could distinguish between "Jeff demonstrated this in a video" and "this is the documented procedure."
I also built tooling to protect institutional knowledge the org didn't fully control - nightly backups of vendor project history, automated syncs of source repositories into AANP-owned infrastructure, and a React-based chat interface for querying the whole knowledge base. Everything was transferred to the organization before I wrapped up.
The engagement ran about two months. But the real deliverable was something harder to put on an invoice: the confidence that when the next person has a question about why something works the way it does, there's an answer waiting for them.