Notes from the workbench.

From a Jev video to a local decision API

A Two Minute Papers video made the idea click: many model calls should be typed decisions, not chat. I tried Kev, then built local-jev, a tiny local package that reads option logits directly instead of parsing generated JSON.

Local AIramazan-yavuz.tr Read article →

Nobody wrote it. Nobody read it.

Since companies started putting AI into everyday work, messages around me have become longer and less personal. I often spend more time finding the request inside them than answering it.

AI & Workplaceramazan-yavuz.tr Read article →

Chiron: what the strong model leaves behind

For a short while I had access to Anthropic's strongest model. Its working discipline was the part I wanted to keep, so I put those habits into an installable kit for the weaker models that stay. Then I measured whether the kit did anything.

AI & Toolingramazan-yavuz.tr Read article →

Are language models just glorified T9?

I keep saying modern AI is basically fancy predictive text. So I went looking for the evidence, for and against my own claim, and let it change my mind where it should. A non-technical, fully-sourced look at where the comparison is fair and where it falls apart.

AI · Non-technicalramazan-yavuz.tr Read article →

The token meter: why AI agents are rewarded for wasting tokens

You pay coding agents by the token. Output costs about five times more than input, the training rewards length, and every file the agent leaves behind is re-read on the next turn. All three incentives favour more output.

AI & Economicsramazan-yavuz.tr Read article →

Two agents, one channel

Building agent-doublethink, a tool that lets two coding agents on two machines coordinate over an end-to-end-encrypted channel: the one decision about what not to build, an assumption I was sure of and corrected, and the bugs only a live test could find.

AI & Toolingramazan-yavuz.tr Read article →

doublethink, told in commits

How the broker actually got built, read straight off its git history: the over-engineered design I deleted, a packaging decision reversed, the Redis swap, and a live demo you can run in your browser. Every commit hash links to the real diff.

Securityramazan-yavuz.tr Read article →

The web revival: the internet didn't die, I just wasn't on it

A twelve-minute video sent me down a rabbit hole into thousands of hand-made personal sites, the XXIIVV webring, and the Merveilles community: a quieter, more intentional internet that never went anywhere.

The Webramazan-yavuz.tr Read article →

doublethink: ntfy you can trust with private data

ntfy is wonderful because it asks nothing of you. That is also why its topics are effectively public. The design path to keeping the ease while adding genuinely private channels, including a dead end I deleted.

Securityramazan-yavuz.tr Read article →

Making Claude Code talk back

How I built claude-can-speak: the TTS engine bake-off, why naturalness beat multilingual, the two-modes design, and shipping it on npm instead of a Debian package.

Local AIramazan-yavuz.tr Read article →

Retrieval-Augmented Generation: what RAG is and how it actually works

A patient walk through what happens when you bolt a search index onto a language model, why that fixes some problems and not others.

Local AIramazan-yavuz.tr Read article →

lillycoder: a local-first coder REPL with permission gates

A small Python CLI that turns any OpenAI-compatible endpoint into a permissioned coding agent.

Local AIramazan-yavuz.tr Read article →

What the quantization names actually mean

If you have ever stared at Q4_K_M, Q5_K_S, IQ3_XXS and wondered which to pick: this is the practical breakdown.

Local AIramazan-yavuz.tr Read article →

Why clauding was built: enabling real workflows for AI agents without compromising control

Three frontier coding agents, one container, no cross-contamination. The boring engineering that makes it useful.

AI & DevOpsramazan-yavuz.tr Read article →

How a stubborn AI agent led me to redesign how APIs talk to machines

An RFC 9457 extension born from one frustrating debugging session with an autonomous agent that refused to learn.

AI & API Designramazan-yavuz.tr Read article →

The ThinkPad that changed my mind about local LLMs

A skeptical look at a Strix Halo ThinkPad with unified memory. I bought it with low expectations and have been overwhelmed by how much I like it.

Hardware Reviewramazan-yavuz.tr Read article →

A quote at the top of every terminal

The idea was small enough to look easy. The implementation, like a lot of small ideas, turned out to need more care than expected.

Linux & UXramazan-yavuz.tr Read article →

An emoji that gossips about your machine

A small tray toy I built on purpose. It turned out to require more careful engineering than most of my serious tools.

Linux & UIramazan-yavuz.tr Read article →

Parking a laptop battery properly

My laptop lives plugged in. Charge thresholds were not enough, so I wrote a small daemon that uses a kernel feature most tools ignore.

Linux & Hardwareramazan-yavuz.tr Read article →

Running local LLMs without the rituals

A small wrapper around llama.cpp that I wrote because I kept rebuilding the same shell scripts, and one I think is finally worth keeping.

Local AIramazan-yavuz.tr Read article →

Text adventures with an AI parser

An interactive fiction engine where the LLM is the parser, never the storyteller. The line matters.

Game Engine & AIramazan-yavuz.tr Read article →

Substance over polish: a guide to the DACH Lebenslauf

A practical, evidence-based guide to the Lebenslauf in the DACH market: what recruiters actually read in the first seconds, the conventions that quietly sink an application, and how to build one that survives both the human and the parser.

Career & Hiringramazan-yavuz.tr Read article →

A protocol for building better software projects with AI

The Narrative Development Protocol: a practical compass for AI-native software builders.

AI & Softwareramazan-yavuz.tr Read article →

From vision to structure: how projects accumulate limits

Why constraints stack up over a project's life and how to plan for it.

Project Managementramazan-yavuz.tr Read article →

The Narrative Management Protocol (NMP) 1.0

A formal spec for managing the story a software project tells itself.

Protocol Designramazan-yavuz.tr Read article →

Your ideal Linux, Docker, and Node.js development setup on Windows

A practical WSL2-based dev environment that has lasted years.

Dev SetupLinkedIn Read on LinkedIn →

Automated intent classification: let tiny llamas do the heavy lifting

When a fine-tuned 7B model beats GPT-4 on the boring routing problem you actually need solved.

AI & LLMsLinkedIn Read on LinkedIn →

Benchmarking sovereign stack: Llama 3, Phi-3, and .NET 8 in production

Numbers from a real production deployment, not a marketing benchmark.

AI BenchmarkingLinkedIn Read on LinkedIn →

Have you ever tried software development on your smartphone?

A short experiment that became a permanent backup setup.

Mobile DevLinkedIn Read on LinkedIn →

Building fast Docker development workflows using mounted source folders

Bind mounts, hot reload, and the small tricks that make Docker feel native.

Docker & DevOpsLinkedIn Read on LinkedIn →

Crying in Angular

A frustration log from working with Angular's tooling in 2021.

Frontendramazan-yavuz.tr Read article →

Docker build fails because it has no internet

Diagnosing and fixing a containerised build that cannot reach the network.

Dockerramazan-yavuz.tr Read article →

Host an Angular web app in WSL2 and access it from the Windows host

A working WSL2 + VS Code setup for Angular development.

Dev Setupramazan-yavuz.tr Read article →

Materialize.css and parallax

Getting parallax sections to behave properly with Materialize.css.

Frontendramazan-yavuz.tr Read article →

Fix "Unsupported platform for fsevents" in npm

A short fix for the macOS-only fsevents warning when running npm install on Linux or Windows.

Node.jsramazan-yavuz.tr Read article →

Open a Windows port to your local network

Configuring Windows Defender Firewall to expose a development port to the LAN.

Windowsramazan-yavuz.tr Read article →

WSL2 Docker not talking to Windows host

Solving the networking gap between Docker inside WSL2 and the Windows host.

Dockerramazan-yavuz.tr Read article →

Enable running PowerShell scripts on Windows 10

Setting the right execution policy for local script development.

Windowsramazan-yavuz.tr Read article →

Start WSL 2 service on host startup

A small autostart trick for WSL 2 on Windows 10.

Windowsramazan-yavuz.tr Read article →

Get Docker running in WSL 2 independently of Docker Desktop

Running Docker inside WSL 2 without the Docker Desktop license.

Dockerramazan-yavuz.tr Read article →

Sticky!

A first post on the blog. Mostly a marker to confirm the WordPress install was working.

Miscramazan-yavuz.tr Read article →