AI-powered IVR platform for emergency dispatch and crisis management. Originally led development in a 4-person team; independently extended with Gemini integration, model retraining, CI/CD pipeline, and production deployment. My contributions: - Trained custom DistilBERT intent classifier on 50k+ annotated emergency transcripts — 94.2% accuracy, 8 classes, exported to ONNX for 15ms CPU inference - Built CNN emotion detector on RAVDESS audio data — 87.3% accuracy, continuous panic score output - Designed hybrid severity scoring engine (panic 40% + keywords 30% + incident priority 20% + urgency 10%) - Built multi-tenant FastAPI backend with row-level data isolation, WebSocket dashboard, and Celery background processing - Deployed to GCP Cloud Run with GitHub Actions CI and Prometheus observability Stack: Python, FastAPI, Whisper, DistilBERT, ONNX Runtime, Celery, Redis, PostgreSQL, Docker, GCP Live demo: redline-ai-359883234654.us-central1.run.app/demo GitHub: github.com/mangod12/redline-AI