Benedict.

←Back to series

SeriesComplete5 of 5 articles

Building BENEDICT.EXE

How I turned a static portfolio into a fast, grounded conversational system—and what broke along the way.

  1. 01Why I Wanted My Portfolio To Talk BackMay 8, 2026 · 8 min readA normal portfolio is optimised for scanning. I wanted something optimised for asking.
  2. 02The First Rule of AI Latency Is Measuring the WaitMay 9, 2026 · 10 min readMy chat worked, but it felt broken. The fix started with measuring the wait users actually felt.
  3. 03Correct Tool, Wrong ScaleMay 10, 2026 · 10 min readQdrant was not a bad tool. It was just the wrong scale for a small curated portfolio corpus.
  4. 04When Search Is Not EnoughMay 11, 2026 · 8 min readFast retrieval found candidate sources. The next problem was deciding which sources should travel together.
  5. 05The Hardest Chat Bugs Were Software BugsMay 12, 2026 · 8 min readI thought the model was failing. Most of the time, the harness around it was.