Fix: pre-trained models + .gitignore

Fix: lazy load - no training on startup

Fix: lighter models to avoid timeout on free CPU

Initial deployment