https://kevinnie2003.github.io/#writing Notes on making LLM output reliable: validation pipelines, grounding, repair loops, and the bugs that hide behind them. en https://kevinnie2003.github.io/writing/knowing-what-the-button-is-called.html https://kevinnie2003.github.io/writing/knowing-what-the-button-is-called.html Tue, 15 Sep 2026 00:00:00 +0000 Turning QA test cases into scripts a real phone will run. Why the hard part was never the code, how mining the game source beat scanning the device, and what a false pass looks like. https://kevinnie2003.github.io/writing/vibes-with-extra-steps.html https://kevinnie2003.github.io/writing/vibes-with-extra-steps.html Wed, 10 Jun 2026 00:00:00 +0000 Making "is this AI-generated card balanced?" a question with a measurable answer, by calibrating a power budget against real Slay the Spire cards. https://kevinnie2003.github.io/writing/four-views-one-object.html https://kevinnie2003.github.io/writing/four-views-one-object.html Mon, 20 Apr 2026 00:00:00 +0000 A multi-agent loop that turns one reference image into four consistent views. Why the judge is a separate agent, when to redo one view versus all four, and when to stop.