I'm an AI engineer working mostly with TypeScript and Python. Automatone is where I publish clear, practical references for the kind of problems engineers actually search for — an error to fix, a tool to set up, a paper to understand.
The aim is simple: when you land here from a search, you should find a direct, correct answer — not a personal essay wrapped around it. Every piece is researched against primary sources and fact-checked before it ships, so the details hold up.
Topics range from AI/ML explainers to Linux and dev troubleshooting, plus first-person build logs from my own experiments — like porting a CUDA-only image-to-3D model to Apple Silicon. If a claim isn't verifiable, it doesn't go in.
How I test and verify
- One real machine. Hands-on tests run on my own Apple M5 MacBook Pro with 32GB of unified memory. Posts say when results won't carry over to other hardware.
- Versions and dates. Measured posts list the exact software versions and when the run happened, for example Ollama 0.33.3, or Python 3.15.0 measured on 2026-10-13.
- Raw output, quoted as-is. Terminal output in reproduced and measured posts is copied from the session logs of the actual run. Any edit, such as a shortened path, is stated.
- Reproduced or sourced, labelled. Where a post reproduces things itself, each one is marked as reproduced or not, and what I couldn't test is backed by a source instead. The ModuleNotFoundError guide reproduces five of its six causes and says which one it didn't.
- Primary sources. Specs and claims link to official docs, PEPs, and papers. Model details are read from the model files themselves where possible.
- Limits and mistakes stay in. Measured posts state what the test doesn't show. The long-context test documents five ways my first runs produced wrong numbers, and says it didn't measure KV-cache quantization. A follow-up did.
- Open lab code. The TRELLIS.2 port to Apple Silicon, with its benchmark scripts, is public at github.com/sanchez-kim/trellis-silicon.
What I write about
How LLMs are trained and run, built from scratch in plain Python and measured on one Mac: the KV cache, attention, tokenizers, sampling, retrieval, LoRA and QLoRA fine-tuning, quantization, and reward methods such as RLVR. Paper reads link to the original papers.
Fixes for specific errors, including Python's ModuleNotFoundError and SSL certificate failures, Docker running out of disk space, and ComfyUI install and out-of-memory problems. Each one starts from the exact error message and shows how to tell which cause you have.
Setup walkthroughs for running AI and automation tools yourself: local LLMs with Ollama and Open WebUI, installing ComfyUI, Qwen-Image on an Apple Silicon Mac, n8n on a VPS with Docker Compose, and the OpenClaw assistant.
Side-by-side comparisons of local LLM runtimes, AI coding assistants, LLM API and cloud GPU pricing, vector databases, and Python package managers. Prices and specs are sourced and carry the date they were checked; a couple of comparisons are measured on my own Mac.
Hands-on tests and project logs: porting TRELLIS.2 to Apple Silicon, measuring local LLMs at long context and with a quantized KV cache on a 32GB M5 MacBook Pro, and benchmarking image models. Versions, commands, and failed runs are included.
Current stack
How these are written
- →Every claim, number, and version is researched against primary sources and linked.
- →Drafts pass a fact-check pass before publishing — outdated or unverifiable details get cut.
- →Code and commands are minimal and copy-paste ready.
- →No fabricated experience and no SEO filler — just the answer you came for.