Welcome to Herd Mentality

Serving a 70B model on rented GPU since I can't run that on my MacBook Pro, then breaking it on purpose and measuring recovery.

A working glossary of LLM inference workloads running on GPUs

au-radar is a RADAR-consistent benchmark of how well AI describes and reaches Australian federal services. It scores Australia 5.32, and finds that a robots-respecting agent is blocked at the free legal database but not at the silent government register.

ClauseKit runs five bodies of law through an LLM extraction pipeline. 171 of 262 extracted rules never produce a yes or no, and that gap is the result.

Porting a fine-tuned sentence embedding model into a Drupal 10 module, running ONNX inference in-process via PHP FFI. No Ollama, no external API, no vector DB.

A practitioner's mapping of the DTA's mandatory agentic AI controls to real Bedrock Guardrails configuration, tested against a live guardrail.

A tool to search across Commonwealth legislation, for Innovation Month 2026

A decompose-then-verify pipeline that checks each claim in an AI output against its source document, and an honest look at what the RAGTruth numbers actually say about it.

Creating a neural network using base R

Extending Bayesian Networks to model time series