Work
Evals
- tithedoes instruction tuning trade calibration for truthfulness?
- rayleighhow many benchmark items it takes to tell two LLMs apart
- mendelhow many seeds an RL comparison needs
RL and environments
- weebleRL-trained robots you can shove, in the browser
- coldstartseconds from a cold Python start to a working policy
Inference and kernels
- quartetTriton matmul kernels and autotuning across four GPUs
- temperamentwhat FP8 attention costs on an H100
- abacusa from-scratch int8 LLM decoder and SIMD kernels for a laptop CPU
- doomformerneural networks compiled into DOOM maps, where voodoo dolls do the arithmetic
In my free time I contribute to open source: over 80 merged pull requests so far, to PyTorch, vLLM, SGLang, torchtitan, CuPy, Ray, ONNX, Hugging Face’s sentence-transformers, timm and PEFT, and Farama’s RL environments.
Research
- secure inferenceGPU function-secret-sharing keys for transformer-scale models can pass 100 GB; my chunked key generation cuts peak key memory 128× for 1.8× the time
- dealerless keysRing-LPN correlation generators remove the trusted dealer from a linear layer: two GPUs make keys the unchanged evaluator accepts, with a conditional semi-honest proof
- 5-HT2Bmolecular dynamics of a serotonin receptor for Alzheimer’s drug discovery, and which signaling pathway a ligand favors
I still think a lot about cryptography research. If you do too, write to me.
Background
- M.S. Artificial Intelligence, Illinois Tech 2026
- B.S. Computer Science and Statistics, Illinois Tech 2025