I build small tools that do one thing and never ask for an account.

A running engineering log. Deep dives, post-mortems that name the real root cause, and 90 browser tools.

The tools are the part people use. All of them run client-side, so there is nothing to sign up for and nothing to upload. Paste a prompt into the tokenizer and it never leaves the tab. Drop a .gguf file into the inspector and it reads the header locally instead of pulling 40GB of weights first.

The ones I reach for most: GGUF Model Inspector (read a model's guts before committing to the download), GPU VRAM Calculator (will it fit, at which quant), Llama.cpp / Ollama Config Builder (hardware in, flags and a Modelfile out), Prompt Cache Optimizer (including the cases where caching loses money), RAG Chunk Inspector (what your splitter actually produced, which is rarely what you assumed).

gekro.com/news publishes a cited AI briefing daily, generated and pushed with nobody watching. 82 of them since June 2nd. Two misses in that stretch. Both were my bugs and both have a write-up.

Offline dictation for Windows and macOS. Push to talk, sub-second latency, Whisper doing the work, no cloud or subscription or telemetry. I picked up PinW's whisper-key-local and kept building on it because I liked Wispr Flow and disliked paying monthly for something my laptop already does.

Happy to talk about local inference, what an agent loop actually costs per turn, or why your chunker is lying to you.