MindFlow · As of October 2026
AI on the iPhone.
Development, training and evaluation.
MindFlow is my unreleased conversational AI prototype for personal concerns. I develop the native app, fine-tune the language model and evaluate its responses. Answers are generated offline on an iPhone 15 Pro using Qwen3-4B, LoRA and MLX. The examples show what improves, what breaks elsewhere and where the model reaches its limits.
The name describes my aim: a conversation that keeps flowing. This is a development goal, not a demonstrated psychological effect.
How an answer takes shape
Four parts of my work. Open the one you want to explore.
THE CHALLENGING LEARNING OBJECTIVES
What was trained?
Changing the form of address is not the same as better understanding.
du→Sie
The German address form.Here, the focus is on meaning.
Understand meaning and conversation history
What was said — and what holds now?Explore↗Ask relevant questions
Ask a question — or deliberately don’t?Explore↗Continue meaningfully
What does the response add to the conversation?Explore↗Respond without assumptions
What is in the conversation — and what isn’t?Explore↗How was it trained?
Original learning examples. Different learning methods. Controlled comparisons.