Latest in AI & LLMs
Building with AI and large language models from Node.js — coding agents, LLM APIs, tokens and cost, prompts, and the tooling around them.
Jev, TypeSafe's "System One" model, tested in Node.js: 400 support tickets, accuracy, latency, cost and calibration
Jev doesn't generate text: you send a state and typed questions, and it returns typed answers with probabilities. 400 real banking support messages were routed with TypeSafe's Node.js SDK: 92.8% accuracy, a 267 ms median, about $0.02 per 1,000 decisions, and confidence scores honest enough to decide which tickets to automate. Compared with a small local LLM, plus what the SDK's TypeScript types catch and where Jev gets it wrong.
Caveman: make your AI coding agent stop rambling (and cut output tokens ~65%)
caveman is a tiny, MIT-licensed skill that installs into Claude Code, Cursor, Gemini, and 30+ other agents and rewrites their prose into terse, fragment-style answers — the project measures a ~65% average cut in output tokens while leaving your code, commands, and error messages byte-for-byte untouched. One curl | bash, Node ≥18, no telemetry. Here's why it's great.