Live: safevoice-vert.vercel.app
🧪 Status: Closed Beta · NGO-Partner-Pilot · Mai 2026 Live deployt, anonym im Browser nutzbar, im Test mit Tutor + Team und ersten NGO-Partnern. Klassifikator gegen 35 reale Test-Cases validiert. 12 deutsche Paragraphen (11 StGB + NetzDG) durch das Schema garantiert. Noch kein Produktivbetrieb für Massenanwendung — Datenschutzerklärung & Impressum sind Vorab-Versionen, nicht anwaltlich geprüft. Postanschrift gemäß § 5 TMG wird vor Live-Gang ergänzt. Wir suchen Trägerschaft (z. B. HateAid). Feedback und Pull Requests sehr willkommen.
Document digital harassment. Classify under German law. Generate court-ready Strafanzeige in 30 seconds.
SafeVoice ships an autonomous Court-Prep Agent as of May 2026. It is the first of 8 agents planned across the SafeVoice / GitLaw / PMM ecosystem and runs on a native agent-loop (no LangChain — see AGENT_MASTERPLAN.md for the reasoning).
| Court-Prep Agent | POST /agent/court-prep/{case_id} — produces Strafanzeige PDF + NetzDG .eml per platform + Staatsanwaltschaft routing + § 77b StGB Strafantrags-Frist-Check + auto re-archived evidence URLs + § 68 Abs. 2, 3 StPO Anonymisierungs-Antrag when needed. 8 tools, ~30 seconds, replaces ~3 hours of manual work. |
| Agent runtime | backend/app/services/agent_loop.py · court_prep_agent.py · court_prep_tools.py |
| Safety guards (built in from day 1) | max_iterations=10 · max_cost_usd=0.50 · idempotency key (run_id, tool, sha256(input)) · tool-call audit table (agent_runs + tool_calls) · human-in-loop checkpoint before every external send |
| Eval set | evals/agent_court_prep.json — runs on every prompt change |
| Cross-project plan | docs/AGENT_MASTERPLAN.md — 8 agents across SafeVoice, GitLaw Citizen, GitLaw Pro, PMM. #1 in production, #2-3 next. |
The same agent_loop contract is being ported to TypeScript for the GitLaw repo so both projects share one Agent-Runtime pattern.
Anonymous-first (no account needed). UI in DE + EN; the OpenAI gpt-4o-mini classifier handles Turkish and Arabic input text out of the box — UI translations for TR/AR are on the roadmap. DSGVO-by-design: no personal data leaves your browser unless you explicitly submit it.
Every 3 minutes someone in Germany is harassed online. 90% goes unreported because:
- Victims don't know which laws apply
- Evidence disappears (posts get deleted)
- Filing a Strafanzeige takes hours of paperwork
PASTE → CLASSIFY → DOCUMENT → STRAFANZEIGE
10s 3s instant 1 click
Paste a hateful message, link, or screenshot → AI classifies it under German criminal law → evidence is preserved with SHA-256 hash chain → court-ready PDF Strafanzeige with executive summary, evidence exhibits, and AI legal assessment ready for the police.
- Paste text — copy a hateful comment from anywhere, paste it. Most common path.
- Paste a URL — public news/blog URLs work directly. Instagram/X/TikTok/Facebook block server-side scraping by design — for those, screenshot is the reliable path and is also DSGVO-cleaner since SafeVoice never touches the platform.
- Upload a screenshot — OCR is Vision-only via OpenAI Vision (gpt-4o-mini). One codepath in dev and prod, no system-package dependency. Vision returns structured JSON: extracted text + sender_handle + platform_hint, so the @username and platform auto-fill from the screenshot.
Single-tier LLM classifier — OpenAI gpt-4o-mini with Pydantic Structured Outputs. Detects 16 offense categories including death_threat, volksverhetzung, doxxing, intimate_images, sexual_harassment, stalking. Returns severity, applicable German paragraphs, bilingual summary (DE + EN), and a requires_immediate_action flag.
Schema enforcement is server-side. Categories and laws are exhaustive Python enums, so the model cannot invent a category or a paragraph. If the model refuses on safety grounds or the response fails schema validation, the API returns a clean 503 — no silent fallback by design, because a weak classification (e.g. regex misreading a death threat) would be worse than no classification.
Few-shot prompt strategy with explicit doxxing example, idiom false-positive ("Das bringt mich um" → low) and obfuscation case ("Stirbt endlich, du H*re" → critical).
Every /analyze/* response carries an LLMMetadata block — model, prompt/completion/total tokens, estimated USD cost, request_id, and the prompt version that produced it. Same telemetry is persisted to an llm_usage audit table per request via a central llm_gateway module, so per-user / per-feature cost dashboards read from one source of truth. Prompt revisions are versioned (PROMPT_VERSION = "v2") and stored on every classification row, so historical classifications stay attributable when the prompt changes.
Once a case has multiple pieces of evidence, a second AI pass produces a Juristische Gesamteinschätzung: free-text assessment, escalation risk + reason, strongest charges with strength scores (strong / medium / weak), and prioritised next steps with deadlines. Auto-refreshes when evidence or victim_context changes. Embedded in the Strafanzeige PDF so the police get one document.
- SHA-256 content hash (sha256:…) per evidence
- UTC timestamp with timezone
- archive.org backup link when scraping
- Browser-side hash verifier — anyone (police, court, lawyer) can paste the original text into SafeVoice and confirm it produces the same hash. Web Crypto API, no network.
A4, court-ready, ~8 KB:
- Executive summary card at the top — complainant, severity badge, case-id, incident counts (with critical count highlighted), statutes — readable in 3 seconds
- Förmliche Strafanzeige body with the victim's personal data substituted in (no [PLACEHOLDER] markers if a name is provided)
- Beweismittel as numbered exhibits — each with severity-colored dot, italic quote with thin left rule, classification line, statutes, mono SHA-256 footer
- KI-gestützte Bewertung as the last section — typographically distinct (tracked-out caps label + italic disclaimer) so police instantly see this is the AI's voice, not the victim's
- Page footer with case-id and page numbers, 30 mm left margin for 2-hole punching
- Tiny
✂fold mark for envelope folding
Plus alternate templates: NetzDG-Meldung (24 h / 7 d platform deadline), allgemeiner Bericht.
3-step Onlinewache flow built into the case page: select Bundesland → copy the prepared text → open the form and paste. Or download the .eml for Apple Mail / Outlook / Thunderbird with attachments pre-attached. PLZ-driven recipient pre-selection.
cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload # :8000cd frontend
npm install
npm run dev # :5173- Frontend: http://localhost:5173
- API docs: http://localhost:8000/docs
OPENAI_API_KEY=sk-… # required — without it /analyze returns 503
DATABASE_URL=… # optional, defaults to sqlite:///./safevoice.db
Single platform: Vercel (Frankfurt, Fluid Compute).
| Layer | Where |
|---|---|
| Frontend (Vite SPA) | Vercel static |
| Backend (FastAPI) | Vercel Python Function via api/proxy.py |
| Database | Neon Postgres (Vercel Marketplace integration) |
| OCR | OpenAI Vision (gpt-4o-mini) — single codepath |
| Classifier + Legal AI | OpenAI gpt-4o-mini |
vercel.json rewrites /api/(.*) → /api/proxy?_p=$1; the ASGI shim in api/proxy.py re-applies the original path before delegating to the FastAPI app whose routers are mounted at both / and /api/.
Env vars on Vercel: OPENAI_API_KEY, DATABASE_URL (auto via Neon), CORS_ORIGINS, VITE_OPERATOR_NAME, VITE_OPERATOR_EMAIL, VITE_OPERATOR_CITY (Impressum, § 5 TMG).
# Analyze (backend/app/routers/analyze.py)
POST /analyze/ingest — text → classify (no DB write)
POST /analyze/url — scrape URL → classify (returns 422 + helpful hint for IG/X/TikTok)
POST /analyze/case — case-level RAG analysis
POST /analyze/chat — case-aware follow-up Q&A
# Cases (backend/app/routers/cases.py)
GET /cases/ — list
GET /cases/{id} — detail + evidence
POST /cases/ — create
PUT /cases/{id} — update title + victim_context
DELETE /cases/{id} — cascade delete
POST /cases/{id}/evidence — append evidence (auto-syncs from frontend)
# Reports (backend/app/routers/reports.py)
GET /reports/{id}?report_type=… — preview text
GET /reports/{id}/pdf?… — A4 PDF (Strafanzeige | NetzDG | general)
— accepts ?victim_name=…&victim_address=…&victim_phone=…&victim_email=…
GET /reports/{id}/legal-pdf — standalone legal-analysis PDF
GET /reports/{id}/court-package — ZIP bundle
POST /reports/{id}/eml — RFC 5322 .eml export with attachments
# Legal AI (backend/app/routers/legal.py)
GET /legal/{case_id} — case-level analysis (auto-cached server-side)
# Upload (backend/app/routers/upload.py)
POST /upload/screenshot — OCR via OpenAI Vision (gpt-4o-mini)
# Auth (optional, NGO/lawyer flow) (backend/app/routers/auth.py)
POST /auth/login — magic link
POST /auth/verify
GET /auth/me · PUT · DELETE
GET /auth/me/export — Art. 20 data export
# Health
GET /health · /api/health
Full interactive OpenAPI docs: http://localhost:8000/docs (or /api/docs on Vercel preview).
| Paragraph | Offense | Max Penalty |
|---|---|---|
| § 130 StGB | Volksverhetzung (Incitement to hatred) | 5 years |
| § 185 StGB | Beleidigung (Insult) | 1 year |
| § 186 StGB | Üble Nachrede (Defamation) | 1 year |
| § 187 StGB | Verleumdung (Slander) | 5 years |
| § 201a StGB | Intimate image violation / Deepfakes | 2 years |
| § 238 StGB | Nachstellung (Stalking) | 3 years (5 with aggravating factors) |
| § 241 StGB | Bedrohung (Threat) | 2 years |
| § 126a StGB | Gefährdende Verbreitung personenbezogener Daten · Strafbare Bedrohung | 3 years |
| § 263 StGB | Betrug (Fraud) | 5 years |
| § 263a StGB | Computerbetrug (Computer fraud) | 5 years |
| § 269 StGB | Fälschung beweiserheblicher Daten | 5 years |
| NetzDG § 3 | Platform removal obligation | €50M fine |
12 entries total — exhaustive list, enforced server-side via Pydantic enum (model literally cannot return anything else).
| Layer | Technology |
|---|---|
| Frontend | React 18 + TypeScript + Vite + Tailwind 4 |
| Backend | Python 3.12 + FastAPI |
| Hosting | Vercel (frontend static + Python Fluid Compute) · Frankfurt |
| Database | Neon Postgres (Vercel Marketplace) |
| AI Classifier | OpenAI gpt-4o-mini (Pydantic structured outputs) |
| OCR | OpenAI Vision (gpt-4o-mini) |
| Evidence | SHA-256 hash chain + browser-side Web-Crypto verifier |
| Reports | ReportLab (A4 PDF) + Python email (RFC 5322 .eml) |
| Auth (optional) | Magic-link, DB-backed sessions |
safevoice/
├── api/proxy.py — Vercel Python Function entry point (ASGI shim)
├── backend/app/
│ ├── main.py — FastAPI app + middleware
│ ├── database.py — SQLAlchemy models + seed
│ ├── models/ — Pydantic domain (Case, Evidence, …)
│ ├── services/
│ │ ├── classifier.py — orchestrator (single-tier LLM)
│ │ ├── classifier_llm_v2.py — gpt-4o-mini + Pydantic .parse()
│ │ ├── llm_gateway.py — central LLM gateway (tokens + cost + request_id)
│ │ ├── legal_ai.py — case-level second-layer analysis
│ │ ├── ocr.py — OpenAI Vision OCR (gpt-4o-mini)
│ │ ├── scraper.py — public web scraper
│ │ ├── evidence.py — SHA-256 hashing
│ │ ├── pdf_generator.py — A4 Strafanzeige PDF (exec summary + exhibits + AI block)
│ │ ├── report_generator.py — Strafanzeige / NetzDG / general body builders
│ │ └── legal_pdf.py — standalone NGO-grade legal PDF
│ └── routers/ — FastAPI routes (analyze, cases, reports, legal, upload, …)
├── frontend/src/
│ ├── pages/ — Home, Analyze, Cases, CaseDetail, Login, Impressum, Datenschutz
│ ├── components/
│ │ ├── ReportModal.tsx — Vorschau + Senden tabs, shared victim form
│ │ ├── SendReport.tsx — recipient picker, .eml builder
│ │ ├── CaseEditor.tsx — add evidence + edit context (tabbed)
│ │ ├── EvidenceCard.tsx — inline @author edit + hash verifier
│ │ ├── HashVerifier.tsx — browser-side SHA-256 verifier
│ │ ├── OnlinewachePanel.tsx — 3-step copy/open/paste flow
│ │ └── … — SeverityBadge, LawCard, AcknowledgementBanner, SafeExit
│ └── i18n/ — DE/EN translations
├── requirements.txt — pinned for Vercel build
├── runtime.txt — python-3.12
├── vercel.json — function config + rewrites
├── schema.dbml — DB schema for dbdiagram.io
└── Dockerfile — production deployment (alternative to Vercel)
Everything beyond the quick start lives in docs/.
docs/DEPLOY.md— deploy guide, env varsdocs/DESIGN.md— design system + product principlesdocs/ROADMAP.md— what's nextdocs/AI_FLOW.md,docs/CASE_CRUD.md,docs/USER_CRUD.md,docs/CLASSIFICATION_API.md— per-topic deep dives
Meeting + presentation artefacts in docs/meeting/ (tutor deck, code walkthrough, study guide, quiz).
SafeVoice provides legal context as general information, not legal advice. For individual advice, contact HateAid or a qualified attorney.
Dieses Projekt ist Teil eines Open-Source-Ökosystems für digitale Demokratie:
| Projekt | Frage | Link |
|---|---|---|
| FairEint | Was sollte Deutschland anders machen? | GitHub · Live |
| GitLaw | Was steht im Gesetz? | GitHub · Live |
| Public Money Mirror | Wohin fließt das Steuergeld? | GitHub |
| SafeVoice | Wer wird online angegriffen? | GitHub |
Alle Projekte: github.com/mikelninh · Unterstützen: Ko-fi · GitHub Sponsors