Amsterdam
Hi, I’m Justin — a product manager who runs his own AI stack at home
By day I work on financial products in Utrecht at Rabobank. In the evenings I run language models on my own hardware and write down what I measure, including the parts that fail.
About Me
I am a product manager in working in Utrecht, living in Amsterdam. Most of my working life is spent deciding what gets built and what does not, which turns out to be reasonable preparation for a hobby that consists largely of finding out whether something works before believing that it does.
At home there is an Intel Arc Pro B70 serving a 35-billion-parameter model, a proxy in front of it, and an agent on top that reads my mail and files my documents. None of it goes near a cloud. What you will find here is the measured version of that: what runs, what broke, and the numbers that made me change my mind.
35B
Parameters served at home
75
Tokens per second, measured
0
Requests that leave the house
What runs here
01
The model
An Intel Arc Pro B70 running Qwen3.6-35B-A3B on llama.cpp, with a LiteLLM proxy in front so that swapping models is a line of configuration.
02
The agent
Hermes, talking to the model through the proxy, with tested wrapper commands instead of letting it invent API calls.
03
The inbox
A small Swift app that reads my mail, pulls out the appointments and deadlines, and proposes them before anything is written to my calendar.
Blog
-
Trilium has an official web clipper for Chrome and Firefox. I use Safari, so I wrote one. It keeps the HTML instead of converting to Markdown, fills the note in from templates, and hands its network requests to a native app because WebKit will not let an extension talk to a server on your own…
-
Everything lands in iCloud Mail and nothing happens to it. A local model reads the inbox and files appointments and deadlines into iCloud Calendar and Reminders, without a copy of my mailbox going anywhere.
-
The rule I ended the first article with turned out to be about vLLM, not about the card. Swapping the runtime made a much larger model run faster.