About

I'm Shawn, a security practitioner and a lifelong learner. For 14 years I've worked across the sharp end of the field — incident response, red teaming, vulnerability discovery and exploit development, and malware reverse engineering — the kind of work that only gets better the more you refuse to stop learning.

Well before LLMs made it obvious, I saw the writing on the wall: AI was going to reshape all of it. Rather than wait, I went and learned the fundamentals of machine learning and AI from the ground up, early enough to understand why these systems behave the way they do, not just how to prompt them.

For the past year I've been pointing that knowledge back at my own craft, applying AI to malware reverse engineering. In practice that means building agents for reverse engineering tasks, fine-tuning models for security workloads, and experimenting with models for classification, byte-level analysis, and other security-specific problems. It's given me an insatiable appetite to learn and build as much as time allows.

Outside work, I spend a lot of time running ML infrastructure in my homelab. I'm interested in inference performance, GPU behavior, agent systems, memory, ML/AI ops, and the engineering needed to make these systems reliable under real workloads.

This blog is where I write about that work: LLM inference and serving, quantization, fine-tuning, agentic systems, GPU performance, and occasional reverse engineering topics. The goal is practical measurement from systems I actually run, not vendor summaries or recycled benchmarks.

Some related code is on GitHub: vLLM serving on the DGX Spark, a kernel guide for consumer GPUs, an inference gateway, and a mechanistic-interpretability puzzle solution. Security tooling stays private for obvious reasons.

How I Write

• Measured where possible. Performance numbers come from hardware I run, with versions, configs, and dates included when they matter.
• Corrections stay visible. If I get something wrong, I'll update the post and keep enough context for the correction to be useful.
• Dates matter. ML infrastructure changes quickly. Posts here are snapshots of what was true at the time, not permanent advice.

If you spot an error, send it my way. The best technical writing gets better when people push on the details.