mimo

mimo takes real conversational turns and fine-tunes a local language model on Apple Silicon (Metal/MPS) to replicate the authentic speaking style, vocabulary, and rhythm of individual speakers.

It integrates with the local fleet:

• paranco lifts chat exports and archives out of macOS protected folders into ordinary directories.
• whatsapp parser directly ingests WhatsApp chat exports (.txt and .zip) for iOS and Android, stripping invisible Unicode markers and formatting multi-line turns.
• LoRA on Metal (MPS) trains low-rank adapter weights on Apple Silicon M4 without sending a single byte off the machine.
• Ollama Adapter compiles and registers calibrated persona models (e.g. eugenio-clone) for zero-latency inference in the terminal and native app.
• Online web sessions are switched off. The page at nerln.github.io/mimo/ no longer collects or sends messages, and nothing goes to the local training corpus from the web.

git clone https://github.com/nerln/mimo
cd mimo
./build.sh                                     # builds Mimo.app
python3 engine/whatsapp_parser.py chat.zip     # imports a WhatsApp export
python3 engine/extract_turns.py                # extracts turns from WhatsApp and Paranco inbox
python3 engine/train_lora.py --speaker Eugenio # trains LoRA on Apple Silicon MPS
python3 engine/ollama_adapter.py --speaker Eugenio --test

The macOS App

A native SwiftUI application (Mimo.app) built with SwiftPM and ad-hoc code signing:

• Storage & Paranco: routes inspection, file lift monitoring, and speaker word count statistics.
• WhatsApp Import: native NSOpenPanel file picker supporting .txt and compressed .zip exports.
• Local Fine-Tuning: speaker selection, base model picker (qwen3:4b-instruct-2507-q4_K_M, olivera:q4), epoch controls, and real-time loss tracking.
• Clone Chat: instant conversational interface with the local persona model.
• Online Sessions: switched off. Room codes, invite links and dataset synchronization from the web no longer work.