How We Built a Sub-200ms Multilingual Chat System Translating 100+ Languages with Our Own LLM

Short summary

A hotel chain built a custom multilingual LLM fine-tuned on hospitality data after commercial translation APIs failed on tone and domain vocabulary. The system translates guest-staff conversations in 100+ languages with sub-200ms latency, running 4–6× faster than commercial APIs across 700+ hotels. The model was trained on 18 months of curated hospitality data including hotel collateral, real conversations, and synthetic dialogue corrected by native speakers.

• •Custom iRoom LLM model outperforms commercial translation APIs on hospitality tasks by 4–6× in speed and accuracy
• •Sub-200ms latency achieved through end-to-end infrastructure control and fine-tuning on domain-specific data (100+ languages)
• •Two-year build process driven by three critical failures of commercial APIs: tone preservation, domain vocabulary, and operational constraints (pricing, rate limits)

Generated with AI, which can make mistakes.