Large language models and AI engineering: agents, retrieval, evaluation, and building systems where the model is only one component.
• views
Zero schema violations in 1,330 Gemini API calls comparing structured output and function calling, but valid records were still wrong.
Read →• views
Trains a segmentation model on a Colab TPU runtime and documents exactly what breaks converting it to LiteRT for an Android edge deployment.
Read →• views
Firebase AI Logic removes the leaked API key, but a real signed-in user can still drain your Gemini budget without App Check and server-side limits.
Read →• views
A Gemini model version bump can silently change what your app refuses, and no normal test suite catches it. Here is a CI gate that does.
Read →• views
A row access policy silently disables BigQuery's vector index pre-filter, so every tenant query turns into search-then-filter that can return zero rows.
Read →• views
A sheet-metal shop retrained MediaPipe Model Maker for gloved hands - it fixed the classifier, but the frozen palm detector still capped end-to-end accuracy.
Read →• views
A support team asked Gemini to box product defects in customer photos - here is what 300 tickets revealed about where its coordinates can be trusted.
Read →• views
A law firm needed to find 'negligence' by meaning, not keywords, while still filtering by date and department - here's the BigQuery ML query that did both.
Read →• views
A custom JAX loss function that occasionally produced NaN gradients quietly corrupted a 128-chip TPU cluster - here's the guard that stopped it, not resharding.
Read →• views
A telecom's support tickets started mentioning '5G outages' their schema didn't recognize. Here's how Gemini's zero-shot classification caught it.
Read →• views
Keras 3 lets you write one model definition and swap between JAX and TensorFlow backends in Agent Platform Workbench without touching your code.
Read →• views
A working TensorFlow model does not need a rewrite for Gemini Enterprise Agent Platform (formerly Vertex AI). Wrap it in a container and deploy.
Read →• views
A loss curve can drop smoothly while a model silently stops learning. Gradient norm monitoring is how you catch it before it costs you.
Read →• views
NLP meets Knowledge Representation to resolve Batman's naming paradox from The Big Bang Theory - a rigorous look at semantic ambiguity.
Read →
