SENTRY: An Anti-Deepfake Speaker Verification Agent

Monolith Modules

• agent/modules/ml_engine: dual-core authenticity scoring (identity + liveness)
• agent/modules/vector_db: Milvus vector vault and behavioral baselines
• agent/modules/streamer: 200ms chunking, Silero VAD, latency anomaly monitoring
• agent/modules/llm_agent: trust-state orchestration and active defense

Legacy per-service folders under services/ are kept as source history/reference, but deployment now runs the single monolith app in agent/main.py.

Quickstart

1. Create local env vars and fill placeholders:
   • cp .env.example .env
   • set MINIO_ROOT_USER, MINIO_ROOT_PASSWORD, and at least one LLM key (GOOGLE_API_KEY or OPENAI_API_KEY).
2. Build and start stack:
   • docker compose up --build
3. Check monolith + module health endpoints:
   • Monolith: http://localhost:8080/api/v1/health
   • ML Engine: http://localhost:8080/api/v1/ml/health
   • Vector DB: http://localhost:8080/api/v1/vector/health
   • Streamer: http://localhost:8080/api/v1/stream/health
   • LLM Agent: http://localhost:8080/api/v1/agent/health

Local API Smoke Tests

• curl -s http://localhost:8080/api/v1/health
• curl -s http://localhost:8080/api/v1/ml/health
• curl -s http://localhost:8080/api/v1/vector/health
• curl -s http://localhost:8080/api/v1/stream/health
• curl -s http://localhost:8080/api/v1/agent/health

Architecture

See docs/ztcv_architecture.md.

GitHub Safety Checklist

• Keep .env untracked (already in .gitignore) and only commit .env.example.
• Never commit real credentials; use placeholders in tracked files.
• If any real key was ever used locally during development, rotate it before making the repository public.