ai_playground
A from-scratch LLaMA-style transformer used as a lab bench: grouped-query attention, RoPE, RMSNorm, SwiGLU, KV cache — plus mixed-precision training, distributed setups, and GPU performance work. Every notebook opens in Colab.
Greg Leclercq
@ggreg·Mountain View, CA·_:: 0x80 ::_Software engineer at Roblox, working on reliable and efficient systems. Lately: LLM internals and training performance, agent architectures, and RL for infrastructure. Longer-standing interests in compilers, distributed systems, and the layers close to the metal.
Current work
ai-deep-dive
A coding agent built from first principles — no agent framework — together with the model it runs on. The implementation, benchmarks, exercises, and the book that explains them all share one source of truth.
atlas
An RL-based Kubernetes autoscaler implementing the AWARE paper (USENIX ATC 2023). DDPG for continuous horizontal and vertical scaling, meta-learning for workload shifts, and a bootstrap scheduler so it can be rolled out safely.
zigvoy
Learn Zig by building an L7 reverse proxy. Seven test-driven exercises; when the last one goes green you have rate limiting, circuit breaking, health checks, and metrics in a working proxy.
Selected earlier work
kaleidoscope
An alternative take on the LLVM tutorial in C++ — the same toy-language compiler, but with the engineering practice the original leaves out: unit tests, a real build, and a style you could grow a language on.
python-synapse
Distributed communication for Python built around nodes and actors, abstracting the transport underneath. ZeroMQ for the wire, JSON-RPC for the messages, and a pub/sub announce service to find everyone else.
libsimplenet
A small C network library on top of libev — event-driven sockets without the ceremony, for when you want the loop but not the framework.