41 billion tokens later: dogfooding local inference routing

85 engineers, 41 billion tokens, 97.5% served locally and Claude bills 46% lower: what we learned dogfooding PaletteAI Inference Launchpad for a month.