AI Compilers, Systems & Agents

I'm Patrick Toulme

Software engineer at Google in New York on the TPU XLA and Gemini RSI teams. I work across the TPU stack — the TPU backend compiler and XLA, TPU performance optimization, JAX/Pallas compilation and performance, and Gemini TPU, serving, and inference optimization — along with Gemini RSI, agent, and agentic harness research. I also created and open sourced pyptx, a Python DSL for writing NVIDIA PTX kernels on Hopper and Blackwell, callable from JAX and PyTorch, and built AgentSage, a replayable record of how coding agents build software, and HarnessGym, a loop that makes coding agents generate and reuse their own tooling. Previously on the MTIA compiler team at Meta and at AWS Neuron working on the Trainium compiler and NKI kernels. I write about AI compilers, AI systems, and JAX/PyTorch on my blog JustAByte.

Location New York, NY
Education Georgia Tech MS, UVA BA
Patrick Toulme

Where I've Worked

2026 — Present

Software Engineer

Google — TPU XLA Team · Gemini RSI Team

TPU backend compiler and XLA, TPU performance optimization, JAX/Pallas compilation and performance. Gemini TPU, serving, and inference optimization. Gemini RSI, agent, and agentic harness research.

XLA TPU JAX Pallas Gemini RSI Inference Agents
2025 — 2026

Compiler Engineer

Meta — MTIA Compiler Performance Team

Worked on GenAI inference compilation in vLLM and high performance inference compilation from FX IR to MTIA ISA. Also worked on the bringup of new MTIA silicon.

PyTorch FX IR MLIR Custom Silicon
2023 — 2025

Machine Learning Engineer II

Amazon AWS Neuron — Annapurna Labs

I led the bringup of the JAX backend for Trainium. I also made GSPMD work on Trainium and work performantly. Worked on native codegeneration and the NKI compiler. I also worked on the bringup of Trainium2. I also wrote the first collective matmul for Trainium — open sourced here.

JAX XLA PJRT GSPMD NKI
2022 — 2023

Software Engineer

Amazon AWS AI Bedrock — Titan Model Training

Worked on the early Bedrock training service. Trained very large LLMs on GPUs.

PyTorch NeMo Distributed Training A100

Things I've Built From Scratch

A Python DSL for writing NVIDIA PTX kernels, an empirical record of how coding agents build software, a self-improving harness loop for coding agents, and a custom tensor processing unit — from RTL hardware to an MLIR compiler and PJRT runtime — that runs JAX and executes Llama as a single fused megakernel.

pyptx

Python DSL → NVIDIA PTX (Hopper + Blackwell)

A Python DSL where the function body is the PTX instruction stream — one PTX instruction = one Python call. Direct Hopper + Blackwell ISA support: wgmma, TMA, tcgen05, mbarriers, cluster barriers. Callable from JAX (via typed XLA FFI) and PyTorch (eager + torch.compile + a C++ extension fast path). Includes maintained GEMM, RMSNorm, SwiGLU, and grouped GEMM kernels for both architectures, plus a PTX → Python transpiler that round-trips byte-identical on 218+ real-world kernels (CUTLASS, Triton, DeepGEMM, ThunderKittens, fast.cu). 815 TFLOPS bf16 GEMM on H100 (beats cuBLAS at ≥6K), 1314 TFLOPS on B200.

Python PTX Hopper Blackwell JAX PyTorch DSL Compiler

AgentSage

A video player for code — replayable agent sessions

The empirical record of how agents build software. AgentSage captures every step of a coding-agent session — file writes, shell commands, reasoning, errors, subagent calls — and replays it as a scrubbable timelapse you can cite moment-by-moment. Arena matches run multiple agents on the same prompt on a shared timeline, and an Elo leaderboard ranks agents and models on evidence rather than opinion. Open-source capture client, CLI, and a searchable archive of recorded sessions.

Agents Coding Agents Evals Elo CLI Open Source

HarnessGym

Coding agents that build their own tooling

Run a coding agent on a hard task, generate the reusable tooling it was missing, then replay the next fresh session with that tooling activated. HarnessGym is a controlled Attempt → Reflect → Build → Qualify → Replay loop: generated artifacts are qualified in a fresh workspace (failures quarantined, not propagated), every MCP tool call is logged to verify the tooling is actually used, and an optimization mode scores workspaces and keeps the best checkpoint. Runner-agnostic (Codex, Claude Code, offline fake runner); core is pure Python stdlib.

Agentic Harness MCP Coding Agents Python Open Source

Custom TPU

JAX → PJRT Runtime → MLIR Compiler → RTL

A fully custom tensor processing unit built from the ground up: Verilog RTL hardware, an MLIR-based compiler (JAX → HLO → MLIR → ASM → VLIW → Binary), and a PJRT runtime that plugs directly into JAX. No CUDA, no hand-written kernels — pure compiler codegen. The compiler fuses entire models into single megakernel binaries, and it now runs Llama end-to-end.

Verilog MLIR PJRT JAX VLIW Llama

Academic Background

Georgia Institute of Technology

M.S. Computer Science — Artificial Intelligence
Graduated December 2025

University of Virginia

B.A. Computer Science
Graduated May 2022

Thomas Jefferson High School

Science and Technology
Graduated May 2018

Research

🏆 Best Paper Award

Marcus: A Chatbot for Depression Screening Based on the PHQ-9 Assessment

A comprehensive study contrasting the effectiveness of screenings by "Marcus," a BERT-based chatbot, against traditional PHQ-9 assessments. Developed a prototype application integrating BERT for linguistic analysis with DialogFlow and Kommunicate APIs.

ACHI 2023 — The Sixteenth International Conference on Advances in Computer-Human Interactions

Technical Skills

Languages & Compilers

  • Python, C++
  • PyTorch, JAX
  • XLA, GSPMD, PJRT
  • MLIR, LLVM
  • FX Graph IR

Hardware & Kernels

  • MTIA (Meta)
  • GPU, TPU
  • Neuron Core (AWS)
  • Triton, Pallas
  • NKI

Machine Learning

  • Transformer Models
  • Distributed Training
  • LLM Pretraining
  • Mixture of Experts
  • RLHF

Agents & Inference

  • Agentic Harnesses
  • Coding Agents & Evals
  • MCP
  • LLM Serving
  • Inference Optimization