07 · How I built this

Nobody likes building a portfolio. So I built mine with agents, from thirteen months of evidence.

Building a portfolio usually means staring at a blank page, trying to remember what you did. I wanted a story I could stand behind, so I treated it like an AI product: where did each claim come from, what would change my mind, and who approves it? Digging up the evidence made the story more specific, and a lot more fun.

13 monthsof scattered work, turned into one story I can stand behind

My roleOperator and reviewer
WhenSep – Oct 2026
ToolsCursor agent, MCP servers, browser automation, my own intake pipeline
Outputthis site, a resume, a positioning strategy

GenAI is changing the world fast, and I wanted to understand where I fit in at such an exciting moment: what the landscape looks like, and which roles match my experience and skills. The jobs had new names: AI enablement lead, agentic product manager, technical deployment lead. My day-to-day requires me to be flexible, so my work is scattered across chat threads, tickets, wiki pages, calendar invites, a Git history, and document libraries. I needed to turn all of it into one cohesive story.

I was the operator and the reviewer. An agent drafted the plan and I edited it twice before anything ran. Agents did the mining; I approved every risky step on an approval card and made the final call on every claim.

Let the evidence change the story. A claims ledger tagged every prior resume claim as verified, flagged, estimate, or retire, and the mining had to confirm, correct, or remove each one. For approvals I chose a hybrid policy: approve the risky writes and let the reads run. That gave the agent room to work without giving it the final word.

  1. six systemstickets, chat, calendar, Git, documents, meetings
  2. agents mineMCP servers, and browser automation where there's no API
  3. claims ledgerverified, flagged, estimate, or retire
  4. I approveapproval cards on every risky write
  5. this siteplus a resume and a positioning strategy

The claims ledger is open source now, rebuilt with made-up data; the intake method is next.

Claims ledgerfour rows from my real run
The claim, as first draftedTagWhat the evidence said
Cut a four-month intake process to about a weekverifiedThe ticket and the meeting recording agree.
Classified 275 AI requestscorrectedThe board had 279 by February, so the site says 279.
Classification accuracy of 99%estimateTrue in testing, not in production. Say "in our testing."
Grew the opportunity pipeline 256%retired256 was the size of the board, not growth.
No tag went public until I approved it.I approve
The ledger kept the story honest: one claim confirmed, one corrected, one softened to "in our testing," and one retired.
Built withMCP clientsPlaywrightClaims ledgerHuman approvalsWeekly digest automation

A story I can stand behind. The result wasn't a pile of facts. It was a cohesive story, a weekly digest of what's happening in the market, a clear sense of how to position myself, and a real picture of the roles out there. I understand the landscape a lot better now, and I have a direction for what could come next in my career.

Every number on this site had to earn its place. The ledger checked each claim against a system of record before it went public. One headline number didn't survive, and several got more specific.

I do, and it keeps changing. The method is open source, rebuilt with made-up data, so anyone can run it without seeing anyone's employer. The real run stays private.

  • Systems of record beat memory. I rediscovered work I'd forgotten about that made a real impact.
  • Approval fatigue is real. Approve the risky writes, and let the reads run.
  • It's never finished. This is a reflection of me and my story, so I keep iterating on it. Getting really loopy!