Most voice note applications rely on third-party cloud APIs (OpenAI, AssemblyAI, Anthropic) for transcription and summarization. This introduces monthly API subscription costs, latency overhead, and privacy risks as personal thoughts and meeting recordings leave the device.
Haven AI is an open-source, voice-first productivity application where 100% of the AI processing happens on-device. No API keys, zero server costs, and total data privacy — functioning even in complete airplane mode.
On-Device AI Pipeline
Audio Capture → whisper.rn Transcription → Qwen 2.5 0.5B (llama.rn) JSON Extraction → Natural Language Date Parsing → On-Device Semantic Search (Cosine Similarity)
Model Optimization & Native Inference
Running LLMs on mobile devices requires aggressive quantization and memory management. Haven AI utilizes Qwen 2.5 0.5B Instruct quantized to Q4_K_M (~400MB) via llama.rn (C++ llama.cpp bindings) and whisper.rn for speech-to-text.
- Automated Note Intelligence: After recording, Whisper transcribes audio locally. The transcript feeds into Qwen 2.5, returning structured JSON with a concise summary, bullet points, and actionable task items.
- Natural Language Reminder Parsing: Extracted tasks like "call accountant tomorrow at 5pm" are parsed by a zero-dependency date parser and scheduled as native notifications via
expo-notifications. - Vector Embeddings & Semantic Search: Note content is embedded locally into floating-point vectors. An AI chat interface computes cosine similarity over stored note vectors directly on 4GB RAM devices, providing context-aware RAG retrieval.
Key Technical Achievements
- 100% private, on-device AI voice note pipeline with zero cloud API dependency.
- Under 5-second transcription and JSON extraction using Qwen 2.5 0.5B via llama.rn.
- On-device vector semantic search running locally on 4GB RAM mobile devices.