4 things I've built — what still works, and what's been left alone.
A personal investment research assistant for hobby investors
Tell it a company, sector or theme and it gathers evidence from public sources (prices, news, filings, ASX announcements), then an AI turns that into reports where every claim is cited and checkable. An optional thesis layer tracks a long-horizon idea as evidence builds for and against it. It never trades or tells you what to buy. Python, SQLite and any OpenAI-compatible model.
Token-efficient orchestration layer for the OpenCode coding agent
An OpenCode plugin plus a benchmark harness that aims to get the most completed coding work out of each scarce premium-model token: a GLM model does the everyday work and ChatGPT is kept for the hard parts. Every optimisation has to earn its place with benchmark evidence — tool-output filtering passed, output discipline did not and was removed.