Joel Reymont
AI Infrastructure & High-Performance Architecture · GenAI / RAG / Agentic Systems · MLIR / OpenXLA · CUDA / Triton / PTX · Ex-CTO/CEO
Schwyz, Switzerland · Remote or on-site in Switzerland · corp-to-corp · [email protected] · github.com/joelreymont · hemis.dev · linkedin.com/in/joelreymont
Summary
AI infrastructure architect and compiler / low-level systems engineer with 20+ years of experience, building high-performance AI systems end-to-end — from GPU/TPU kernels and compilers up to production GenAI platforms.
- High-performance AI architecture: custom MLIR compiler stacks, OpenXLA integration, and GPU/TPU kernel tuning (CUDA, Triton, JAX), backed by executive leadership (former CTO/CEO).
- Advanced GenAI & system design: production-grade platforms with code-knowledge graphs, multi-source RAG, and hybrid semantic-vector search (pgvector, BM25) over complex enterprise codebases.
- Agentic workflows & AI adoption: autonomous work-item → PR agents and low-code, chat-driven application builders that lift developer productivity.
Equally deep in IR / VM / JIT codegen, distributed systems, and firmware / reverse engineering. Learns fast, ships complete solutions, works independently.
Core skills
- Compilers & codegen: OpenXLA / XLA, JAX, MLIR/LLVM-adjacent toolchains; IR / VM / JIT design; DSL design; Ghidra Sleigh → Binary Ninja transpilers (pixie); OCaml compiler (DWARF, debugger); LLDB plugin; a Common Lisp compiler (mamushi).
- GPU & accelerators: CUDA 12.6 (nvcc, sm_87), custom Triton kernels, NVIDIA PTX, Nsight Compute / Systems profiling, TensorRT, JAX & PyTorch on GPU, Jetson Orin.
- Languages: Rust, Zig, C, C++, OCaml, Common Lisp, Forth, Objective-C, Erlang/Elixir, Go, Julia, Python.
- AI-driven engineering: Agentic workflows with Claude Code / Codex / OpenCode; built agent tooling (Banjo, Dots, pz, Mnemos); verified, test-driven model-generated delivery.
- Systems & tooling: Fault-tolerant distributed systems (Erlang/BEAM); embedded Linux (Yocto), ARM, FreeRTOS; reverse engineering (IDA Pro, Binary Ninja, Ghidra); Nix, Bazel; macOS internals & device drivers.
Experience
Compiler Engineer — Hemis (own company)
habu: Checked Forth → PTX kernel DSL & AI tooling · Oct 2025 – Present
- Building habu, a self-hosting, statically checked Forth — native macOS ARM64 and Linux AArch64 engines that JIT-compile to machine code, emit AOT signed binaries, and rebuild byte-for-byte. Hindley–Milner type inference with first-class row variables over both the data and return stacks; the compiler verifies every definition's declared stack effect (979 words self-certified).
- Active work: a tile-level GPU kernel DSL that emits NVIDIA PTX (ptxas, sm_87 / Jetson Orin), whose checker makes classes of GPU bugs — wrong memory space, out-of-bounds access, divergent collectives — unrepresentable at the source level; aimed at reliable LLM-generated kernels.
- GPU kernels (Odin): real-time 4-camera edge perception on Jetson Orin — CUDA 12.6, custom Triton kernels, JAX / PyTorch / TensorRT; profiling/tuning with Nsight Compute / Systems (occupancy, memory throughput, warp stalls, register pressure).
- AI-driven engineering tools (open source, Zig): Dots (120★ task tracker for AI agents), pz (90★ minimal coding agent), Banjo (42★ Claude Code adapter for Zed), Mnemos (second brain for code); built procedures for fully test-verified, model-generated delivery.
- Compiler/toolchain contributions: DWARF debug info + bytecode-debugger features for the OCaml compiler (upstream PR #14353); zdb LLDB type-formatter plugin for Zig; zcheck property testing.
Compiler / Rust Engineer — Elodin (OpenXLA / JAX)
Aug 2025 – Oct 2025
- Upgraded the simulation core to the latest XLA / JAX stack; built a cross-platform subset of XLA with Nix and Bazel, packaged as a static library for safe in-process embedding (eliminating cross-memory-space context-lifetime bugs). Modernized the Nix build infrastructure; contributed to the Bevy-based simulator.
Compiler Engineer — Sleigh → Binary Ninja architecture compiler (commercial)
pixie / dixie · Oct 2024 – Aug 2025
- A compiler translating Ghidra Sleigh processor specifications into Binary Ninja architecture plugins — DSL parsing, IR, code generation/optimization, and execution on a purpose-built VM (formerly a JIT backend). Delivered across three language targets: C++ (IDA Pro) → OCaml → Rust (Binary Ninja). Adjacent: DWARF and bytecode-debugger contributions to the OCaml compiler.
Firmware / Low-Level Systems Engineer — Drone Forensics (self-employed)
Jan 2022 – Oct 2024
- Built an onboard autonomy computer for DJI Enterprise drones; reverse-engineered Mavic 3 / 3E firmware, the Payload SDK, and the RTK module at machine-code level (ARM, FreeRTOS) using IDA Pro, Binary Ninja, and Ghidra; built a Yocto embedded-Linux distribution.
Founder/CEO — Stegos AG · CTO — Æternity
2017 – 2022
- Founded Stegos; architected a blockchain (Lisp → Rust), raised $15M, and led a multidisciplinary team to delivery. As CTO at Æternity, built and ran a remote Erlang engineering team end-to-end.
Senior Software Engineer — Performance & Distributed Systems (self-employed)
2004 – 2017
- Performance: sped up Firefox startup on macOS via DTrace profiling; ported the High-Performance Erlang compiler (HiPE) to macOS; Erlang NIF disk backend for CouchDB; scaled an Erlang broadcast server to 100k+ users on EC2 at minimal latency.
- Distributed systems & compilers: OpenPoker, a fault-tolerant Erlang poker server sold to Electronic Arts (EA WSOP); a Go XMPP server for Thomson Reuters Eikon; a SQL-like → Erlang compiler in OCaml. Drivers: CoreAudio HAL multichannel driver; LabView USB/GPIB driver (reverse-engineered via IDA Pro).
Selected open source
dots (120★), pz (90★), banjo (42★), habu (typed Forth → PTX), mamushi (Common Lisp), zdb (Zig LLDB formatters), zcheck (property testing), and mnemos — all at github.com/joelreymont. OCaml compiler upstream PR #14353.
Availability
Available immediately for remote contract work, corp-to-corp, or on-site in Switzerland — short or long-term. Spanish citizen (full EU work authorization); Swiss C-permit. No visa sponsorship required.
If you're facing a hard technical problem, I can help. Email me and we'll find a solution.