Every four years, my friend group runs a World Cup pool. The format is simple: everyone gets randomly allocated a bunch of countries, and you earn points as your countries progress through the tournament. The more games your countries win, the more points you accumulate, until eventually someone’s Argentina or France makes it to the final and that person becomes insufferable about it for the next four years. This year I decided to build a website for it.

I am not a sports analyst. I don’t have insider knowledge of team sheet dynamics, injury recovery timelines, or the subtle psychological impact of playing a Thursday night game in Canberra in July. What I do have is Claude, a Python environment, and a dangerous amount of free time. The result is nrl-predictor.ohare.id.au — an AI-powered NRL match predictor that publishes its picks every week and, crucially, publishes its accuracy record even when it’s wrong.

I recently worked through a tutorial using Google’s Agent Development Kit (ADK), a Python framework for building multi-agent AI workflows. The tutorial used a fun premise: an AI pipeline that researches historical figures and generates full movie pitches, complete with plot outlines, box office projections, and casting suggestions. What is Google ADK? The ADK lets you compose multiple AI agents into pipelines using three core building blocks: SequentialAgent — runs sub-agents one after another, passing state between them LoopAgent — repeatedly runs sub-agents until a termination condition is met (using exit_loop) ParallelAgent — runs sub-agents concurrently and merges their outputs Agents communicate through a shared state dictionary, which tools can read and write via a ToolContext object.

I built the following Passphrase Generator using Claude.ai, following Harper Read’s article: My LLM codegen workflow atm.