Chess Personality Project

A chess stylometry system that analyzes playing styles and matches players to historical chess masters.

Overview

This project implements the chess stylometry research from the papers in papers/. We build a system that:

• Parses PGN files and extracts chess games
• Uses CNN + Transformer architecture to create player embeddings
• Compares uploaded games against historical chess masters
• Provides personality matching results via web interface

Note: We focus on stylometry/player comparisons only (not the Maia-2 engine extensions).

Environment

• Use uv for Python environment management
• Run tests with: uv run pytest

Architecture

• Backend: FastAPI with SQLite database
• ML Pipeline: CNN feature extraction � Transformer aggregation � Player embeddings
• Frontend: Vanilla HTML/CSS/JS
• Processing: Async job queue for analysis tasks

TODO:

• Implement vectorization for moves
  • 2 FEN go in (before/after), 1 vector comes out
  • 8x8x17
    • First 12 channels: one 8x8 for each piece color/type
      • zeroes(), 1 for each position holding that piece
    • Next 4 channels: all ones() if player can castle
      • White kingside, white queenside, black kingside, black queenside
    • Last channel: all ones() if player to move is white
  • One "move" is a tuple of two position vectors (before/after)
• Create PGN cleanser to remove games with <10 turns, insufficient metadata, etc.