Contact me Patrick Gawron.
AI Automation Architect
Agentic Workflows · Local-First AI · Secure Business Logic
I build AI agents that run entirely under your control or on your own hardware. They automate your workflow without sending sensitive data to the cloud.
Currently Running daily AI experiments, orchestrating agents, shipping consulting projects. Studying for n8n KI Automation Specialist.
Notes
63 notes. Experiments I ran, news worth a comment, tips from X, and random observations.
Patrick Gawron @patrickgawron · #63
Tuning Qwen3.6-35B-NVFP4 on a single RTX 5090 — vLLM sweet spot
22 configs, one MoE model (Qwen3.6-35B-A3B-NVFP4), one RTX 5090. Sweet spot: 96k context, fp8 KV cache, 2 parallel → 160 tok/s. Full write-up in the article.
Patrick Gawron @patrickgawron · #62
Why Local LLMs Matter in 2026
The case for running models on your own hardware, even when the cloud is cheap.
Patrick Gawron @patrickgawron · #61
Whisper Large-v3 vs Parakeet
Parakeet 1.1B beats Whisper Large-v3 on WER and speed for English long-form transcription.
Patrick Gawron @patrickgawron · #60
The Best Agent Is a Boring Agent
Why the most reliable production agents look surprisingly unambitious.
Patrick Gawron @patrickgawron · #59
Local agent benchmarks - which model survives a real loop
Five local models against the same multi-step agent task on RTX 4090. Qwen 3.5 14B wins on tool-call reliability.
Patrick Gawron @patrickgawron · #58
Stop Writing README Files for Claude
Why CLAUDE.md should be short, terse, and scoped.
Also
Not everything I do is AI.
Photography
Composition, sharpness, bokeh. Mostly product and detail shots. Sony Alpha, manual glass.
Creative tech
Laser cutting, UV printing, acrylic photo gifts. Physical things from digital designs.
Home lab
Minisforum server, Synology NAS, UniFi 10G, RTX GPU running local models 24/7.