Play 57 — AI Translation Engine
Enterprise translation engine — two-layer architecture: Azure Translator for bulk NMT (130+ languages, custom glossary enforcement) + GPT-4o LLM post-editing for nuanced content (marketing, legal, medical), quality scoring (BLEU/COMET + LLM judge), HTML/markdown preservation, batch document translation with progress tracking.
Architecture
Full architecture details:
architecture.md
How It Differs from Related Plays
| Aspect | Play 49 (Creative AI) | Play 57 (Translation Engine) |
|---|---|---|
| Input | Campaign brief (monolingual) | Source text in any language |
| Output | Creative content (1 language) | Translated text in 130+ languages |
| Quality | Brand adherence | BLEU/COMET translation quality |
| Glossary | Brand voice rules | Domain terminology enforcement |
| Cost model | Per campaign | Per 1K words translated |
| Bulk | N/A | Batch document translation |
Key Metrics
| Metric | Target | Description |
|---|---|---|
| BLEU (general) | > 40 | N-gram overlap with reference |
| COMET | > 0.80 | Neural quality estimation |
| Glossary Compliance | > 95% | Terms translated per glossary |
| Product Name Preservation | 100% | Product names never translated |
| Markup Preservation | 100% | HTML/markdown tags intact |
| Cost per 1K Words | < $0.20 | NMT + LLM refinement |
Cost Estimate
| Service | Dev | Prod | Enterprise |
|---|---|---|---|
| Azure OpenAI | $60 | $500 | $2,000 |
| Azure AI Translator | $0 | $200 | $800 |
| Cosmos DB | $5 | $120 | $500 |
| Azure CDN | $5 | $50 | $200 |
| Azure Blob Storage | $3 | $25 | $80 |
| Key Vault | $1 | $3 | $10 |
| Application Insights | $0 | $25 | $80 |
| Total | $74 | $923 | $3,670 |
Detailed breakdown with SKUs and optimization tips:
cost.json· Azure Pricing Calculator
WAF Alignment
| Pillar | Implementation |
|---|---|
| Reliability | Azure Translator SLA, glossary enforcement, quality thresholds |
| Cost Optimization | NMT for bulk ($0.05/1K), LLM only for marketing/legal/medical |
| Performance Efficiency | >5000 wpm basic, batch document translation |
| Operational Excellence | BLEU/COMET tracking, glossary auto-update, batch checkpoints |
| Security | Key Vault for API keys, no PII in translation logs |
| Responsible AI | Quality scoring with human review triggers, cultural adaptation |
FAI Manifest
| Field | Value |
|---|---|
| Play | 57-ai-translation-engine |
| Version | 1.0.0 |
| Knowledge | F1-GenAI-Foundations, F2-LLM-Selection, T2-Responsible-AI, T3-Production-Patterns |
| WAF Pillars | performance-efficiency, cost-optimization, reliability, responsible-ai |
| Groundedness | ≥ 85% |
| Safety | 0 violations max |
Last updated on