09 · Open source

How I built this portfolio, rebuilt so you can run it.

My real evidence lives in systems nobody outside my company can open. So I rebuilt the method with a made-up person, a made-up company, and six made-up systems. You can watch every claim get confirmed, corrected, or retired, and see exactly where the method still needs a person.

6 of 7claims tagged the way a careful reviewer would; the one miss is published

My roleDesigned and wrote it, alone
WhenOct 2026
DataSynthetic only; MIT licensed
StatusWorking; eval published

Most people write a resume from memory, then go looking for proof. I did the reverse for this portfolio, and it changed the story: one headline number didn't survive, and several got more specific. But the real run can't be shared. It's built on my work history, and that stays private.

I designed and wrote all of it, from scratch: no employer code, prompts, data, or names. The demo follows a fictional enablement lead, Aria Chen, at a fictional company. Every claim, record, and number in it is invented.

Changing a claim's public status is a write, so a person approves it. Mining and reading never wait. Tagging a claim verified, or retiring it, waits for someone to read the list and run approve --yes. Without that flag, nothing is written.

  1. draft claimseverything the resume wants to say
  2. mine six systemstickets, chat, calendar, Git, documents, meetings
  3. tag each claimverified, flagged, estimate, or retire, with the quote behind it
  4. a person approvesafter reading every tag
  5. ledger outthe claims you can defend, and why

A read-only MCP server lets an agent inspect the ledger. It can't approve anything.

Built withPython, standard library onlyCLIClaims ledgerHuman approvalRead-only MCP serverEval against a gold setunittest
4
tags, and every claim gets one: verified, flagged, estimate, or retire
1
headline retired: the number was real, but it was a snapshot, not growth
6 of 7
claims tagged the way a careful reviewer would, with the miss printed instead of hidden

A gold set records what a careful reviewer would tag, and the eval prints every miss. The miss is the point: a true fact that lives in only one system, a standing monthly review on a calendar, gets flagged instead of verified. A person would sign off on it. The heuristic won't, and the README says so.

$ python3 ledger.py eval
accuracy  6/7  (0.857)
weak spot C07: got flagged, gold verified

Anyone can. Swap in your own draft claims and exports from your own tools, keep the shape, and run it. The real run behind this site stays private.

Repository: github.com/christine-q-nguyen/portfolio-from-evidence. The story behind it: How I built this portfolio.

  • Single-source facts still need a person. Don't turn 86% into a headline; publish the miss.
  • Approve the risky writes, let the reads run. It's the load-aware version of a human in the loop.