Letter Boxed

Puzzle solver and visualizer for NYT’s Letter Boxed

It started as a Java project for CS 112 in Spring 2022. That summer I rewrote the recursive backtracking solver in TypeScript, built a React frontend to visualize it, and deployed it to GitHub Pages.

The solver did all its work on the client, and heavy puzzles would freeze the UI. I chased that problem for a while: first I moved the algorithm to a Google Cloud Function (which stopped the freezing but added network latency), then rewrote that function in Go to bring latency back down. Eventually I brought the solver home to the client for good — running it in a Web Worker so the heavy computation never blocks the interface.

The current version leans into that client-side approach. The solver tracks a puzzle’s twelve letters as a 12-bit bitmask, so checking whether a word chain uses every letter is a single bitwise comparison; it precomputes letter adjacency and filters the ~26K-word dictionary down to the few hundred words a given puzzle allows before searching. Solve returns the first valid answer, while Find Best searches for the optimal one — fewest words, then shortest total length — offloading that heavier combinatorial search to the GPU via WebGPU compute shaders that extend word chains in parallel across multiple passes, with a CPU fallback. A slider steps through each stage of the solution, and a generator produces random puzzles weighted by English letter frequency.

Tech stack: React, TypeScript, Tailwind v4, shadcn/ui, Zustand, Vite, WebGPU (WGSL), Web Workers, Bun, Vitest.