The design decisions, the dead ends, and the stuff that doesn't fit in a README.
RSS feedPure heuristics miss deep strategy. Pure LLMs are too slow for every turn. Here's why poke-ai splits the decision loop between the two instead of picking one.
Notes from building a ~130-route production platform that real VIT students use instead of VTOP — on why the hard part was never the dashboard.
Why a warehouse rover trained directly on hard layouts never learns anything, and how promoting it through four difficulty tiers at an 80% success threshold fixed that.