Mick Lab
Research centered on generative AI applications
日本語 | English

Mick Lab — where experiments become working apps

Every Mick Lab app is designed, implemented, tested, and published using only an Android phone. Browse the apps by category and open each product page for details.

Browse the apps
Built on Android only

Android-only development stack

All apps, including LLM AI Server with llama.cpp, Dual AI Chat, and Dual AI Reversy, are developed directly on Android without relying on a PC.

See how Userland + Termux + Copilot CLI + custom scripts run GitHub CI →

LLM run & train (Android)

2 apps

On-device GGUF inference server, plus a training-data generation and on-device training tool.

LLM AI Server with llama.cpp icon

LLM AI Server with llama.cpp

Validation tool combining GGUF local inference, vision (Qwen3-VL), GPU backend support, on-device API/WebUI, and MCP integration

 GGUF Generator icon

GGUF Generator

On-device tool that goes from Q/A training-data generation to creating, continued-training, and LoRA fine-tuning small GGUF models

AI apps (Android)

3 apps

End-user apps powered by generative AI.

Dual AI Chat

Chat app for Ollama-compatible APIs, with two AIs conversing.

Dual AI Reversy

Reversi app with independent Human / CPU / AI control for Black and White

 ClarityPDF icon

ClarityPDF

On-device PDF create/edit/convert with OCR / AI-OCR (summary and custom prompts)

Browser / PWA

2 apps

Install-free web apps that run in the browser, offline.

Local LLM (GGUF) inference viewer

Demo that downloads a GGUF from a fixed URL, stores it in IndexedDB, and runs inference in the browser

 VOSK speech-to-text icon

VOSK local speech-to-text

PWA that keeps a VOSK small model on-device and transcribes speech continuously, offline (Japanese/English)

Dev / CI

Guide

How generative AI apps are developed and published using only an Android device.

Development / CI notes

Android development and GitHub integration with Userland, Termux, and Copilot CLI

Latest experiment notes

Recent updates and experiments.

  • GGUF Generator: added the product page, user manual, and bilingual privacy policy (training-data generation, new GGUF creation / continued training, LFM2 LoRA, keyword evaluation)
  • LLM AI Server with llama.cpp: new default vision model Qwen3-VL with automatic context expansion for large images/long prompts (GPU→CPU fallback); smarter model search (model/size/quantization pull-downs, device-suitability rating, direct mmproj download); updated to llama.cpp b10621 (v0.3.0), added /api/diagnostics, and over-limit inputs now return a proper server error
  • ClarityPDF: added the product page, user manual, and bilingual privacy policy (PDF create/edit, convert/compose, OCR / AI-OCR, summary and AI prompts)
  • LLM AI Server with llama.cpp: updated screenshots to the latest JPEGs and fully revised the user manual to match the current app (collapsible UI, status header, Direct Run section, GPU compute backend, mmproj, MAINTAIN MODEL, and more)
  • VOSK local speech-to-text: download a VOSK small model (Japanese/English) into the browser and run continuous on-device, offline recognition; PWA-ready to add to the home screen
  • Local LLM (GGUF) inference viewer: download GGUF and run actual inference inside the browser
  • LLM AI Server with llama.cpp: refreshed the product page, manual, technical spec, and privacy policy (MCP / function calling, multimodal API, WebUI integration)
  • Dual AI Reversy: added the product page, user manual, and bilingual privacy policy
  • Dual AI Chat: added the product page and manual