AI and automation

The GRANT is the only control a tool-calling agent can’t argue with: what it stops, the 11.7-minute scan it doesn’t, and the failure it never reaches.

• MCP ·
• Agents ·
• PostgreSQL

A working demo tells you nothing about your next change. How to build an evaluation set from real questions, score retrieval separately from the answer, calibrate the judge, and gate every prompt change in CI.

• Evaluation ·
• LLMs

Why the Claude Code install line fails in Command Prompt, the four errors that name your shell, and the 599-build gap before WSL 2 exists.

• Agents ·
• LLMs

The difference between a workflow and an agent, costed end to end on an invented system called Lamplighter: 2.4x on a good run, 35x on a bad one, and the error rate that actually decides the build.

• Agents ·
• Architecture

Retrieval changes what a model can see; fine-tuning changes how it answers. A diagnostic for telling which problem you have, and the arithmetic for deciding when a lexical index is worth adding.

• RAG ·
• LLMs

Corpus size decides between retrieval and a long context window. Fine-tuning sits on another axis. The crossover is derived twice; the fine-tune never pays.

• RAG ·
• LLMs ·
• Cost

Why a semantic cache breaks even at a 0.01 per cent repeat rate, how to set the threshold from a false-hit budget, and what the semantic half is worth.

• Caching ·
• Cost ·
• RAG

Why the splitter loses the number and the model never had it: split probability derived at LlamaIndex’s shipped defaults, compounded across a 1,200-table corpus, and what a bigger chunk costs.

• RAG ·
• LLMs

Tokenisation to the KV cache, then the arithmetic: 128 KiB a token on Llama 3.1 8B, 16 GiB of cache for one full context window, and why prompt caching buys 22 per cent of the bill and 7 per cent of the clock.

• LLMs ·
• Cost ·
• Caching