ENHR
Start a Project

Notes From the Workbench

What we learn building AI products, client systems and research - written up with the numbers and the mistakes left in.

Subscribe via RSS

  1. 9 min read    From Blueprint to Archviz Renders in Headless BlenderOne blueprint, one brief and a day of Claude writing Blender scripts: seven 2400×1600 renders and a walkable web model. What it took, and what broke.

  2. 10 min read    How We Laser-Cut Our Homepage Out of Blender RendersThree devices, gold wireframes and scale-model figures, built in Blender by an AI agent and animated without WebGL. Plus the shadow bug that took four attempts.

  3. 5 min read    The Moat Is the Verifier, Not the GeneratorGeneration got cheap. What makes AI output worth shipping is whatever checks it - four of our projects, four verifiers, and the one check a bot couldn't do.

  4. 11 min read    How We Distilled Jev for MOJ KRAJ. Triage for $0.10A 322M model on our own GPU now makes most of MOJ KRAJ.'s triage calls. It learned from $0.10 of a commercial model's answers and ties the production LLM.

  5. 5 min read    A Judge That Mostly Agrees Is Not a JudgeWe tested a small model as MOJ KRAJ.'s fact-check judge. It was faster, agreed on the easy calls, and would have published 15 of 20 articles the real judge held.

  6. 5 min read    One Newsroom, Ten Languages, 1.4M ArticlesHow we translate a Croatian newsroom into ten languages every day - the English pivot, the cost per article, and the bug that duplicated a quarter of our rows.

  7. 5 min read    The Judge Failed Croatian, Not the BotOur eval judge failed a correct answer 12 times in a row. The agent was right; the judge couldn't read Croatian negation. How we tested the tester.

  8. 5 min read    Every Obvious Fix Made Our Croatian Subtitles WorseWe benchmarked Whisper on real Croatian podcasts, then tested every obvious fix. None of them helped - and weeks later we shipped one of them anyway.