Engineering and product updates from Bayes Impact.https://www.bayesimpact.org/https://www.bayesimpact.org/blog/post-turn-classification/https://www.bayesimpact.org/blog/post-turn-classification/We asked the answering model to file its own paperwork through a mandatory tool, and smaller models showed us the cost. Moving that work to a separate call after the reply fixed it.Wed, 30 Sep 2026 00:00:00 GMTJérémie Doucyhttps://www.bayesimpact.org/blog/sovereign-ai-public-services/https://www.bayesimpact.org/blog/sovereign-ai-public-services/Why public institutions run open models like Gemma 4 on their own infrastructure, and how the Bayes Platform keeps citizen data in-house, open-source, and free of lock-in.Thu, 17 Sep 2026 00:00:00 GMTPaul Duanhttps://www.bayesimpact.org/blog/evaluate-ai-agent-before-launch/https://www.bayesimpact.org/blog/evaluate-ai-agent-before-launch/A self-reported accuracy rate is not evidence. How review campaigns on the Bayes Platform let the professionals of a public service validate an agent's answers, blind, before it goes live.Thu, 13 Aug 2026 00:00:00 GMTAnna Bauerhttps://www.bayesimpact.org/blog/tool-calling-open-models/https://www.bayesimpact.org/blog/tool-calling-open-models/We named our RAG tool poorly, and Gemma 4 told us so. Here's what we learned about designing tool schemas for multi-model platforms.Mon, 20 Jul 2026 00:00:00 GMTJérémie Doucy