Personal AI that can stay close to your own data

01

What local-first means in CyanBridge — LiteRT + llama.cpp on phone, Ollama on laptop

• 01Prefer on-phone local inference via curated catalog: LiteRT-LM for Gemma 4 E2B/E4B (mixed 2/4/8-bit) and llama.cpp for Qwen3.5 0.8B gguf (Q4_0), with templates qwen_chat/gemma_it, GPU/CPU backends and NPU coming soon for Snapdragon.
• 02Support services you run on a computer you control — RemoteOpenAiClient to Ollama at http://localhost:11434, llama.cpp server or Model Studio — for larger context or lower phone load, with multilingual TTS chunking locally.
• 03Keep personal dataset preparation (OpenCode/Claude/Codex, WhatsApp, ChatGPT, Takeout) and Obsidian vault Markdown (.md + YAML, Storage Access Framework scoped folder) plus 9 native plugins local by default: Local Agent, Walking Aid (LiteRT vision), Meeting Spark Notes, Live Caption Relay, Hands-Free Translator, Errand Brain, Auto Diary, Auto Audio, Visual Diary.
• 04Make cloud fine-tuning or hosted inference an explicit, separately quoted choice rather than an invisible dependency — local via LiteRT/llama.cpp needs no subscription, hosted plans $1–$5–$20 are optional.

02

Personal datasets without inventing a new data silo — including Obsidian

CyanBridge is designed around exports and files users already control, such as chat histories, coding-agent logs, and other personal archives, plus Obsidian vault Markdown synced via SafKnowledgeRepository (choose existing vault or create plain .md vault, scoped read/write, #tags optional, periodic sync; external .md files remain plaintext and are not encrypted by Memory Vault). The point is to turn user-owned context into a model or retrieval workflow while keeping the data path inspectable.

03

Explore the local-AI stack