# Musings of an AI Wrangler

> I am a software engineer that loves solving complicated problems. Once in a while I solve a problem I want to document, so I'll do that here.

## Posts

- [Speculative Decoding on GLM-5.3-Flash: DFlash vs Native MTP](https://mazurov.dev/posts/glm-speculation-benchmark/llms.txt): A four-arm benchmark of speculative decoding on GLM-5.3-Flash (NVFP4, TP2 over two GB10 nodes): DFlash BF16, DFlash FP8 draft KV, native MTP, and no speculation. Native MTP's bundled draft head costs about 20% of short-prompt throughput and leaves an 89% larger KV pool - a good trade for long-context serving.
- [Making Your Hugo Blog Citable by LLMs](https://mazurov.dev/posts/making-your-blog-citable-by-llms/llms.txt): Practical steps to help LLMs accurately cite your blog: llms.txt, Schema.org BlogPosting markup, structured metadata, and problem-first writing. Implemented on a Hugo site with PaperMod.
- [Fixing the '?' Hostname Problem on OpenWrt Access Points](https://mazurov.dev/posts/openwrt-ap-hostname-sniffer/llms.txt): OpenWrt dumb APs show '?' for client hostnames because /tmp/dhcp.leases is empty. A tcpdump script sniffs DHCP Option 12 from bridge traffic and writes lease entries that LuCI can display.
- [ESPHome YK-H/531E Climate Component: IR Control for AC Units](https://mazurov.dev/posts/esphome-ykh531e-climate-component/llms.txt): Custom ESPHome component for controlling AC units via the YK-H/531E IR remote protocol. Supports temperature control, multiple modes, fan speeds, and swing on ESP32 and ESP8266.
- [Server to Server Authentication with Google Provider in oauth2-proxy](https://mazurov.dev/posts/server-to-server-jwt-with-oauth2-proxy/llms.txt): How to authenticate server-to-server API calls through oauth2-proxy using Google Service Accounts and JWT token exchange.