Projects

• Lumen

  • Statically-typed language and compiler in idiomatic Rust.
  • Full pipeline: lexer, parser, name resolution, type checker, typed HIR, optimizer, codegen.
  • Runs on a stack-based bytecode virtual machine.
  • CFG mid-level IR, C transpiler, and bytecode verifier.
  • Rich diagnostics with error codes and multi-span labels.
  Rust Compiler Type system Bytecode VM
  View on GitHub

• Rust Chess Engine

  • Hand-written engine in safe Rust, no unsafe, no chess crates.
  • Perft-verified legal move generation.
  • Iterative-deepening PVS: transposition table, quiescence, null-move, LMR.
  • Tapered PeSTO piece-square evaluation.
  • UCI interface, PGN/SAN, opening book, and terminal UI.
  Rust Chess engine UCI Search & eval
  View on GitHub

• High Perf. Inference CUDA Kernels

  • Step-by-step optimized CUDA kernels for the core operations of LLM inference.
  • Five ops (GEMV, Softmax, RMSNorm, SiLU, attention decode), each a progression from a naive kernel to an optimized one.
  • Memory-bandwidth bound: every optimization chases the 360 GB/s DRAM bandwidth, not FLOPS.
  • Verified against a double-precision CPU reference across many shapes and edge cases; all 161 checks pass.
  • Written for and measured on an NVIDIA RTX 3060 12GB (Ampere GA106, sm_86).
  CUDA LLM inference GPU kernels Performance
  View on GitHub