A practical guide to hardware, OS, and llama.cpp tuning, built from a year of experiments on a single consumer CUDA workstation.
Long-form Writing
Writing
40Writing, tutorials, and thoughts on tech, tools, and whatever else stays with me long enough to write down.
2026
Local benchmarks comparing Gemma 4 26B under standard, QAT, MTP, and combined QAT+MTP configurations on an RTX 4070.
I burned API credits comparing ChatGPT Images 2.0, Nano Banana, and Imagen on prompts designed to be annoying.
A messy still-frame trailer experiment, and the weirdly practical lessons hiding inside it.
End-to-end Gemma 4 setup with official MTP assistant drafter models on llama.cpp.
End-to-end Qwen3.6-35B-A3B MTP setup on llama.cpp with throughput notes and MTP speculative decoding speedups.
A note-taking app becomes a weird place very quickly when you ask it to be your records system, admin dashboard, and substitute for thinking.
Practical Qwen3.6 setup on 12 GB VRAM using llama.cpp.
A data analyst's mostly futile quest for the perfect workflow.
End-to-end Gemma 4 setup on mainline llama.cpp with text and vision throughput notes.
The tech stack powering this site - Netlify, Eleventy, Alpine, and Tailwind - and all the services that make it tick
A Raspberry Pi 4B, a Pi camera, an HLS stream, and a Cloudflare tunnel - my take on the window-to-the-world idea.
Introducing /folio/ - themed, standalone web pages for things I obsess over. Part deep-dive, part visual essay, part personal reference guide.
aidar tracks stylistic signals associated with AI-era writing across the web
You probably don't need >$100 plans for coding nowadays.
Squeezing every token per second - the sequel
Ramblings on what gets quietly lost when intelligence becomes a utility.
A lazy person's guide to exploratory data analysis that actually produces something at the end
2025
It's irresistibly fun, and yet, something doesn't sit quite right.
2023
2022
Get back HW accelerated playback on Fedora.
How to jump ship your hobby projects.
2021
Pinned
Local LLM Inference Optimization: The Complete Guide
A practical guide to hardware, OS, and llama.cpp tuning, built from a year of experiments on a single consumer CUDA workstation.
Pinned
Optimizing gpt-oss-120b speed on consumer hardware
Squeezing every token per second
Gemma 4 26B QAT + MTP: 100 tok/s Local MoE on 12GB VRAM
Local benchmarks comparing Gemma 4 26B under standard, QAT, MTP, and combined QAT+MTP configurations on an RTX 4070.
Battle of the Image Generators
I burned API credits comparing ChatGPT Images 2.0, Nano Banana, and Imagen on prompts designed to be annoying.
I Tried to Ghiblify a Trailer - Loganime
A messy still-frame trailer experiment, and the weirdly practical lessons hiding inside it.
Running Gemma 4 MTP locally on 12GB VRAM
End-to-end Gemma 4 setup with official MTP assistant drafter models on llama.cpp.
Running Qwen3.6-35B-A3B MTP locally on 12GB VRAM
End-to-end Qwen3.6-35B-A3B MTP setup on llama.cpp with throughput notes and MTP speculative decoding speedups.
What Obsidian Shouldn't Be
A note-taking app becomes a weird place very quickly when you ask it to be your records system, admin dashboard, and substitute for thinking.
Running Qwen3.6-35B-A3B locally with text + vision
Practical Qwen3.6 setup on 12 GB VRAM using llama.cpp.
★ Featured
In Search of the Agent IDE for Data
A data analyst's mostly futile quest for the perfect workflow.
Claude Code is all you need?
I know that "all you need" is annoying but hear me out.
Running Gemma 4 26B-A4B locally on 12GB VRAM
End-to-end Gemma 4 setup on mainline llama.cpp with text and vision throughput notes.
The NEAT Framework: How I Built This Site
The tech stack powering this site - Netlify, Eleventy, Alpine, and Tailwind - and all the services that make it tick
Building My Own WindowSwap
A Raspberry Pi 4B, a Pi camera, an HLS stream, and a Cloudflare tunnel - my take on the window-to-the-world idea.
New Entity: Folio
Introducing /folio/ - themed, standalone web pages for things I obsess over. Part deep-dive, part visual essay, part personal reference guide.
I built a tool to measure how AI-like a website writes. Then I scanned my own.
aidar tracks stylistic signals associated with AI-era writing across the web
Ramblings on Running an Army of Very Smart Five-Year-Olds
truly the year of the agents
Agent-Hopping - A Cost-Effective Way to Access Coding Agents
You probably don't need >$100 plans for coding nowadays.
Running Qwen3-Coder-Next at 40 t/s on consumer hardware (draft)
Squeezing every token per second - the sequel
Don't give away to the gradient descent
Ramblings on what gets quietly lost when intelligence becomes a utility.
Using Claude Code (or any other agent) for data analysis on Databricks
A lazy person's guide to exploratory data analysis that actually produces something at the end
Turning an old webcam + RPi into a camera server
Fun little weekend project
Personal Braumgartner Restoration
Bringing old photos back to life.
GPT‑OSS vs GPT‑5 on GPT‑5’s example prompts
A side-by-side benchmark on web dev tasks
Vault 76 - My Obsidian Vault Structure
How I organize my notes and resources.
When Computers Finally Speak Human
The end of programming as we know it
Ghiblify everything, or should we?...
It's irresistibly fun, and yet, something doesn't sit quite right.
The Unforwarder: How i stopped backing up junk photos
Taming the WhatsApp chaos
Exposing local llm's to the internet (draft)
Your very own AI-API
My favorite meme.
I'm out of blog ideas.
Running your own Copilot in VSCode on a Mac.
With Codellama, Ollama and Continue.
Dumbing down my iPhone.
Trying to stay away from distractions.
Remoting into WSL2 externally - the easy way.
Simple is better than complex.
Repurposing my laptop as a Home Server - Part I
Putting old hardware to good use.
Moving my site from Jekyll to Eleventy - Part II.
What I learned along the way..
Building Mesa (with VA-API) from Source on Fedora.
Get back HW accelerated playback on Fedora.
Moving my site from Jekyll to Eleventy.
How I revamped my personal site.
Migrating Heroku Postgres Database to a Cloud VM.
How to jump ship your hobby projects.
Sync Obsidian notes between iOS and PC.
Take your second brain everywhere with you.
Challenges to AI-ML-based search engines.
Do privacy laws have an impact?