How We Built an AI Translation API with Google Gemini

Why AI Translation? Traditional translation APIs (Google Translate, DeepL) work well for general text, but they struggle with: Arabic dialects — Gulf Arabic vs Egyptian vs Levantine Formality levels — Formal business Arabic vs casual chat Context-aware translation — Technical terms, cultural nuances We decided to use Google Gemini 2.5 Flash as our translation engine. It's fast (< 500ms), understands context, and supports dialect control. Architecture Client → /api/v1/translate → Cache Check (Valkey) ↓ miss Gemini 2.5 Flash ↓ Cache Store (24h TTL) ↓ Response The Prompt The key to good AI translation is the prompt. Here's our approach: Translate the following text from {source} to {target}. Formality: {formality} Dialect: {dialect} Rules: - Maintain the original meaning and tone - Use the specified dialect if provided - Return ONLY the translated text, no explanations Caching Strategy Translation is expensive (API credits + latency). We cache aggressively: Cache key : Hash of text + sour