What I use every day

Everything here runs every day, built and operated by me. No slides: the demos at the AI Practice Day run on these systems.

MitScript

Record meetings, identify speakers, get minutes with corrections and Word export, all on local AI. Speech recognition (Whisper) runs locally, and a local model then cleans up the text. No cloud. It comes with my own Android keyboard for dictation (Mitboard), which uses the same endpoints. So I dictate instead of typing on my smartphone.

In use since: March 2026

Mitboard (illustration): dictation keyboard on an Android smartphone, a dictated sentence appears in a note while the microphone key is active and transcription runs on-device
Illustration, not a live capture

Offtangent

My own open-source agent platform and personal AI assistant. It comes with a companion app and Puck, my own hardware button. It takes in notes, voice messages and tasks and sends them to the right model or agent. It works through longer tasks in the background and shows results on boards. Around 4,400 automated tests check the code. Offtangent runs on my own hardware, and I use it every day.

In use since: March 2026

Offtangent in the browser: strand “Minutes: supplier meeting” with three bullet points, a comparison table of label printers from vendors A, B and C and a recommendation
Real interface, sample content

My own inference hardware

Speech and text models run on-premise on a Mac Studio: Whisper for speech, Qwen and Gemma for text. The demos at the Practice Day run on this machine. There you see what one computer in your own building can do.

KikuchiBot Discovery Interview

An AI-guided needs interview. It asks about your company and your workflows and clarifies your needs before the first conversation with me. The interview is live. There is no public link because usage is limited. I show a demo on request.

Live since October 2026, currently first pilot interviews.

Halfway

Halfway is a side project that I built and run myself. The app finds a fair meeting point in the middle for people in different places. It uses real driving routes instead of a straight line and suggests places to meet there. It runs publicly as a web app and an Android app and is used by people I have never met. That makes Halfway a finished product in real use, not just technology in the background.

Models

I work every day with the large models from Anthropic, OpenAI and Google. Also with open models from China (Qwen, Kimi, GLM) and with models on my own hardware. More important than the model is the question which data may go where. I answer that before the first prompt.

Live at the AI Practice Day

I show MitScript, Offtangent and the inference hardware live at the AI Practice Day.