Why Generic Synthetic Examples Fail in Category-Specific ScenariosA targeted generator-prompt change, what 93.6% actually means, and lessons from AI Waiter4 min read·idirocco.hashnode.dev6d ago
A Small Pipeline, With Explicit Boundaries: Architecture & ImplementationModel choices, orchestration, structured outputs, evaluation, and the production trade-offs I haven't measured yet.6 min read·idirocco.hashnode.devOct 28
A Passing Eval Doesn't Mean Your AI Feature Is ReliableHow I think about evaluation sets, failure analysis, and the gap between a good score and a dependable AI product.8 min read·idirocco.hashnode.devOct 19
My LLM Judge Went From 70% to 96% Agreement With Humans. Here’s What I Learned.What three iterations of an LLM-as-a-judge rubric taught me about evaluation design, regression testing, and synthetic data quality.14 min read·idirocco.hashnode.devOct 15
