SafeVoice 🛡️

Evidence-first AI for documenting digital harassment and preparing a human-reviewable case package.

SafeVoice turns text, URLs and screenshots into a structured record: what happened, what evidence exists, which legal context may be relevant and what a person can review next.

Open the live beta → · Portfolio

Try the workflow

text / URL / screenshot
          ↓
structured extraction
          ↓
classification + legal context
          ↓
evidence hash + source metadata
          ↓
case-level review
          ↓
report / submission draft
          ↓
human review before external action

The goal is not to replace a lawyer, police officer or court. It is to reduce the distance between “this happened” and “I have an organised, inspectable record.”

Proof at a glance

Signal	Current prototype
Evaluation corpus	35 curated cases
Full-pass cases	30 / 35
Severity agreement	94%
Category agreement	89%
Law-set agreement	86%
Forbidden-law false-positive check	100%

These are curated evaluation cases — not 35 real users or validated police cases.

What is implemented

• structured Pydantic outputs rather than unconstrained legal prose
• text, URL and screenshot intake
• SHA-256 evidence hashes, timestamps and source metadata
• browser-side integrity verification with Web Crypto
• multi-evidence case analysis
• reviewable PDF / export workflows
• bounded agent runtime with explicit tools, iteration limits and audit records
• approval checkpoint before any external action

Architecture

React / TypeScript
       ↓
FastAPI + Pydantic
       ↓
structured AI + vision
       ↓
evidence + case services
       ↓
bounded tool loop
       ↓
reviewable output

Stack

Python · FastAPI · Pydantic · React · TypeScript · Postgres · OpenAI structured outputs · Vision · SHA-256 · ReportLab

Why the boundaries matter

SafeVoice can help prepare and organise. It does not determine guilt, provide qualified legal advice or autonomously file consequential legal actions.

Before broader production use, the important next proof is not another feature. It is:

real users → qualified reviewer corrections → regression cases → measurable outcomes

Run locally

# backend
cd backend
pip install -r requirements.txt
uvicorn app.main:app --reload

# frontend
cd ../frontend
npm install
npm run dev

LLM paths require an OPENAI_API_KEY.

---

Built by Michael Ninh in Berlin.