ARC-AGI Solver

An agent for ARC-AGI puzzles — the abstraction-and-reasoning grids where you get a few before/after examples and have to infer the rule, then apply it to a grid you've never seen. Built for Georgia Tech's Knowledge-Based AI course (CS 7637).

The first version was a dispatcher: roughly forty hand-written solvers, each tuned to one problem family, with a front door that guessed which specialist to call. It solved problems and taught nothing general — every new puzzle meant another solver. The rewrite went the other way: nine general transformation primitives, each split into a detect half (does this rule explain the examples?) and an apply half (do it to the test grid), chained so multi-step rules compose out of single-step parts.

The compositional pipeline lands 77% — 74 of 96 — of the course's defined problem set. The dispatcher could only ever be as smart as its longest case statement; the pipeline occasionally solves things I didn't plan for, which is the whole argument for building it that way.

• ~40 HAND-WRITTEN SOLVERS, V1
• 9 TRANSFORM PRIMITIVES, V2
• DETECT / APPLY PIPELINE
• 77% — 74/96 SOLVED