Smartling delivers high-quality translations at scale using Amazon Nova

Serving hundreds of enterprises across more than 450 languages and locales, Smartling provides AI-driven translations. As demand grew, it faced a key challenge: foundational large language models produced fluent translations but often lacked the nuance required to capture tone, terminology, and cultural context. Customers had to spend significant time on manual post-editing to ensure quality and brand consistency.

Olga Beregovaya, vice president of AI at Smartling, says, “Our biggest challenge was striking the right balance between accuracy and fluency. We needed translations that were faithful to our customers’ brand voice while keeping costs manageable.”

Performance was also critical. Latency is a common challenge for large language models, and even short delays can slow down fully automated translation workflows. Smartling needed a solution that could deliver higher-quality results, reduce reliance on human editors, and operate fast enough to keep up with customers’ needs.

As a participant in the AWS ISV Accelerate Program, Smartling decided to expand its use of AWS services to address these challenges. The company integrated the Amazon Nova Pro foundation model within LanguageAI, its AI-powered translation platform, for its balance of quality, cost efficiency, and speed.

To strengthen translation accuracy, Smartling designed its AI workflows with a retrieval-augmented generation (RAG) AI framework on Amazon Aurora that intelligently incorporates translation memory, terminology, and style guides. By collaborating closely with AWS, the team optimized the RAG implementation and reduced purpose-built prompt sizes from 3,000 to just 200 tokens while maintaining translation quality. This optimization enabled the system to process customer-specific requirements like terminology and brand voice efficiently. Combined with a custom prompt-engineering pipeline, the RAG framework helps Smartling deliver high-quality AI translations, with Amazon Simple Storage Service (Amazon S3) storing translation data and Amazon CloudWatch monitoring system performance in near real-time.

A key factor in the project’s success was close collaboration with the Generative AI Innovation Center. Joint benchmarking and optimization helped Smartling reduce computational costs by cutting prompt sizes from thousands of tokens to just a few hundred while preserving translation quality. "The level of collaboration and support from AWS has been outstanding, with Smartling and AWS working as one team to drive innovation and customer success," says Beregovaya. 

By integrating Amazon Nova Pro within LanguageAI, Smartling achieved notable gains in translation quality, efficiency, and cost savings. Benchmarking showed a 26 percent improvement in Bilingual Evaluation Understudy (BLEU) scores, a widely used metric for evaluating machine-generated text by comparing it to human-created reference translations. “Higher BLEU scores mean more accurate capture of tone, terminology, and stylistic preferences, among other criteria,” says Beregovaya. Smartling’s linguists also reported a 30 percent decrease in editing effort, which helped enterprises bring content to market faster.

In addition, Smartling achieved cost savings, with the new solution proving up to 15 times cheaper compared to leading trained machine translation engines, which cost about $80 per million characters. In turn, this helps the company pass savings along to its customers while maintaining high-quality translations. “Amazon Nova has given us the performance, cost efficiency, and quality we need to expand translation at scale. It helps our customers grow globally by bringing content to market faster,” Beregovaya adds.

Alongside these gains, the company also addressed latency challenges. Smartling has not experienced any inference delays, which ensures that translations run reliably in fully automated workflows for its customers. Building on this foundation, Smartling plans to make LanguageAI available through AWS Marketplace and continue refining its performance as adoption grows.