From the desk of a CTO
Notes on building.
Essays on engineering leadership, SaaS architecture, and the strange shape of small technical teams.
How Well Does Qwen3.8-27B Run on the Intel Arc Pro B70?
Qwen3.8-27B on the Intel Arc Pro B70, with benchmark results across llama.cpp and vLLM XPU, including MTP.
Read post2026 is the year of Linux for me: ChatGPT out, AI agents in
A print job vanished with no error message and nothing to search for. One plain-language sentence to an AI agent with terminal access found the cause in minutes. Why I finally adopted Linux.
Read postIntel Arc B70 context decay: the KV cache setting that fixes it
A reader's question led to a depth-swept benchmark on the Intel Arc B70. Generation halves by 32K on stock settings — f16 KV cache on SYCL fixes it.
Read postMigrating Claude Projects to Open WebUI: what the export leaves out, and how to rebuild it
Claude's export doesn't link chats to projects. Here's how I migrated two years of Claude history and rebuilt Projects as Open WebUI Knowledge + Models.
Read postDon't trust your old benchmarks: a weekend of testing on the Intel Arc B70
A Saturday of testing on the Intel Arc B70 produced 8× throughput on Qwen3.6-35B-A3B. Mesa 26.1, backend choice, and ubatch tuning.
Read postKeep your own keys: running a local AI agent on hardware you own
How I run AI agents on my own hardware: a frontier model plans, a local model executes, and my credentials never leave the box. The setup, the risks, the why.
Read postKeep Your Own Data: A Personal Data Sovereignty Stack for 2026
The asymmetry no one's fixing: every AI company has your data and you don't. Here's the personal data sovereignty stack I'm building on an Intel Arc Pro B70 and Qwen3.6-35B-A3B.
Read postHello, again
Kicking off a writing habit on engineering leadership, fractional CTO work, and the strange shape of small technical teams.
Read postRunning Qwen3.6 Locally on the Intel Arc Pro B70: Why I Chose llama.cpp Over vLLM
Real benchmarks running Qwen3.6-35B-A3B on the Intel Arc Pro B70. Why llama.cpp beats vLLM for this model today, and what's coming next.
Read post