Field notes on AI agent observability, GEO and choosing a neutral tooling stack.
2026-06-17
API relays often swap the Claude/GPT you pay for with cheaper models, even prompting them to claim they are still Claude. How behavioral-fingerprint detection works, why asking the model is useless, and the real accuracy.
2026-06-13
A concrete GEO (Generative Engine Optimization) case study: the exact five changes we made to panshi.io/tools so ChatGPT, Perplexity and Google AI Overviews can cite it — with the research behind each.
2026-06-13
Of the 37 LLM tracing/observability tools we track, 29 are OpenTelemetry-native and 8 are not. Here is why that single fact decides your migration cost — and how to tell the difference.
2026-06-13
We self-hosted open-source LLM observability tools and measured the two things that matter: how much instrumentation overhead they add, and how heavy they are to run. Real numbers, honest caveats.