https://blog.lgcode.me/tags/llm/ Recent content in Llm on lgcode Hugo en-us Sat, 29 Aug 2026 09:00:00 -0300 https://blog.lgcode.me/posts/the-cache-that-cancels-your-improvement/ Sat, 29 Aug 2026 09:00:00 -0300 https://blog.lgcode.me/posts/the-cache-that-cancels-your-improvement/ I improved a matching pipeline, measured the improvement, deployed it, and reprocessed the backlog. The backlog reproduced every old decision exactly. The cache was keyed on the input and stored the verdict. https://blog.lgcode.me/posts/the-demo-video-was-the-easy-part/ Sat, 08 Aug 2026 09:00:00 -0300 https://blog.lgcode.me/posts/the-demo-video-was-the-easy-part/ A margin that was arithmetically correct and substantively false, confidence scores from a model that wasn't deployed, and one in five invoice lines failing its own arithmetic. Generating the video took minutes. Making it true took two days. https://blog.lgcode.me/posts/mcp-tool-output-design/ Mon, 03 Aug 2026 10:00:00 -0300 https://blog.lgcode.me/posts/mcp-tool-output-design/ A filter that discarded the answer, a retry that couldn't work, a partial result posing as a total, and cached prices with no freshness signal. All four found in one live session, all four the same class of mistake.