https://pietrolauria.com/ A working lab for applied AI and adtech: notes, projects, project updates and corrections, newest first. en Sun, 04 Oct 2026 00:00:00 +0000 https://pietrolauria.com/projects/ai-analyst https://pietrolauria.com/projects/ai-analyst#2026-10-04-project-update Sun, 04 Oct 2026 00:00:00 +0000 One version passed 18 / 20 in a full live run and went live on 28 September. A newer model also scored 18 / 20, but with two new failures, so under a rule I set before its final run it stays out. https://pietrolauria.com/projects/ai-analyst https://pietrolauria.com/projects/ai-analyst#2026-09-28-project-update Mon, 28 Sep 2026 00:00:00 +0000 The four failed or unrun questions now pass. The advertiser briefings still misstate direction, such as “all campaigns declined” when one rose. One of ten briefing runs passed. https://pietrolauria.com/projects/ai-analyst#corrections https://pietrolauria.com/projects/ai-analyst#2026-09-28-correction Mon, 28 Sep 2026 00:00:00 +0000 The page said the repaired version had not been re-run against the model. A live re-run had started ten minutes before that page went up. https://pietrolauria.com/notes/incremental-compared-with-what https://pietrolauria.com/notes/incremental-compared-with-what#2026-09-27-note Sun, 27 Sep 2026 00:00:00 +0000 Four things get called incrementality in measurement meetings. Two are causal, and one question tells you which. https://pietrolauria.com/projects/ai-analyst https://pietrolauria.com/projects/ai-analyst#2026-09-27-project Sun, 27 Sep 2026 00:00:00 +0000 An assistant for commercial teams: ask about campaign performance or request an advertiser briefing in plain language, and get an answer where Python computes every figure and the model must cite it, so spikes and declines come with evidence you can check.