Sub Rosa: On-Device AI Legal Intake
Sub Rosa records a lawyer's first client meeting and turns it into clear intake notes: facts with the quotes that back them, the standard intake items the client never mentioned, and follow-up questions. Runs on the phone, fully offline. No cloud.
Core trick
Records and diarizes a consultation with Whisper and pyannote, then an on-device LLM on the NPU extracts facts linked to verbatim quotes.
Description
$27
README excerpt
<div align="center"> # Sub Rosa **On-device AI note-taker for a lawyer's first client consultation.** Records → transcribes → extracts structured intake notes — **100% on-device, fully offline.** Built for the **Qualcomm × Meta ExecuTorch Hackathon** (June 27–28, 2026). Target device: **Samsung Galaxy S25 Ultra** · Snapdragon 8 Elite · Hexagon NPU. </div> --- ## What it does Sub Rosa turns a lawyer's first client consultation into clean, structured intake notes — entirely on the phone, with **no network calls at any point**. It records the conversation, transcribes it on-device (Whisper), diarizes the speakers (pyannote, via sherpa-onnx), and runs a small on-device LLM **on the Hexagon NPU** to produce **three outputs**: 1. **Facts** — each extracted fact paired with the **verbatim quote** from the transcript that supports it. 2. **Missing information** — the standard intake items the client did **not** cover, computed as a true **set-difference** against a human-authored, matter-type checklist. *(This is the killer feature.)* 3. **Follow-up prompts** — short, neutral interview questions for the lawyer — never legal conclusions. ### The guiding principle > The model is a **reasoning-over-provided-text engine, never a knowledge source.** It never states the law. All legal scaffolding lives in **human-authored checklists** injected into the prompt at runtime — so a 1B-class model stays accurate and safe with **zero fine-tuning**. The app's legal competence equals