The Future of Agentic Engineering
A five-part series on where software work heads as coding agents become a daily tool. From eighteen years building software and nine across the AI stack: the durable principles that stay fixed, the methods that become executable loops, and the human shift from writing code to steering, verifying, and owning systems of agents.
Building Software Support for a New AI Accelerator
Eleven parts follow a new accelerator from hardware and drivers through PyTorch tensors, kernels, graph compilation, distributed execution, workload runtimes, qualification, and release, then compare the integration boundaries in JAX, MAX, CUDA, ROCm, and Intel XPU.
Fornax - Heterogeneous Sparse-MoE Serving
A series on Fornax, a project to build a Mojo/MAX-native distributed inference engine for placing one large sparse-MoE model across a fleet of commodity machines when the checkpoint exceeds the usable memory of every selected node. It covers accelerator support from hardware to framework APIs, MAX’s compiler and serving stack, the Fornax contracts, and the local kernel patches that made one bounded DeepSeek-V2-Lite path emit tokens on an M3 Max.
Ignis - Finding Your Mojo from DwarfStar
A six-part expedition into Ignis, a Mojo-native agent harness that runs a MAX-served LLM in the same OS process as the control plane. An honest account of what Mojo and MAX actually do at the 1.0-beta frontier - from building a real program on them. (New to Mojo or MAX? Two parts open with from-scratch primers.)
ASR - Notes from Building Speech-to-Text
Field notes from building and tuning automatic speech recognition pipelines - the components around the model and the trade-offs that decide quality in practice. First up: tuning voice-activity detection, using Silero VAD and FunASR’s FSMN-VAD as worked examples.
Welcome
This is my personal blog - where I write up the things I build and what they teach me, usually in more detail than fits in a README.
More writing to come. You can find me on GitHub.