LILT Fine-Tunes Multilingual Generative AI Models with NVIDIA NeMo on AWS

LILT is known for bringing human-powered, technology-assisted translations to global enterprises. Its translation and content generation solution empowers product, marketing, support, ecommerce, and localization teams to deliver exceptional customer experiences to global audiences. To produce content, LILT developed its own multilingual generative AI models using NVIDIA GPUs to power model training, fine-tuning, and retraining. However, as LILT scaled, it needed the ability to fine-tune and retrain models that were five to 50 times bigger—all in real time—and capture higher-quality edits from its linguists.

By adopting NVIDIA NeMo, a cloud-native framework for building generative AI models, LILT has accelerated fine-tuning. NVIDIA NeMo is included as a part of NVIDIA AI Enterprise, a secure, cloud-native, end-to-end software platform that streamlines building and deployment of production-grade AI applications, including generative AI. “The speed and frameworks that NVIDIA provides are so important to LILT,” said Omar Orqueda, vice president, AI research and engineering at LILT. “With these, we’re able to improve our models with the terms, tone, and colloquialisms our customers use—ultimately delivering better multilingual content.”

The combination of NVIDIA’s G4dn.12xlarge GPU instance and NVIDIA NeMo allows LILT to implement a real-time human-in-the-loop approach for all verified translations. This ensures all suggested changes from LILT’s linguists are used for model fine-tuning—helping to produce more accurate content. “The NVIDIA computing power makes our linguists more efficient,” Orqueda said. “They create training data for us that we can use to make our models better.”

As a solution that supports many different use cases, LILT structured its platform to provide customers the flexibility to choose from a variety of models. For multilingual content creation, LILT offers Amazon Bedrock, a fully managed service that makes foundation models from leading AI startups and Amazon available via an API. “We believe in the power of choice for LILT customers, and Amazon Bedrock lets them decide which model is best suited for their needs,” Orqueda said.

The fact that Amazon Bedrock can be deployed on AWS GovCloud (US) also allows LILT’s public sector customers to access a suite of models. “Many of our customers in the public sector need to deploy very sensitive content, so they’re using LILT on AWS GovCloud,” Orqueda said. “That way they can scale their loads and translate content in real time.”

As a solution that supports many different use cases, LILT structured its platform to provide customers the flexibility to choose from a variety of models. For multilingual content creation, LILT offers Amazon Bedrock, a fully managed service that makes foundation models from leading AI startups and Amazon available via an API. “We believe in the power of choice for LILT customers, and Amazon Bedrock lets them decide which model is best suited for their needs,” Orqueda said.

The fact that Amazon Bedrock can be deployed on AWS GovCloud also allows LILT’s public sector customers to access a suite of models. “Many of our customers in the public sector need to deploy very sensitive content, so they’re using LILT on AWS GovCloud,” Orqueda said. “That way they can scale their loads and translate content in real time.”

Not only does LILT allow customers to choose their model, but they can also adapt it based on their own requirements. This ensures high-quality batch machine translations so that customers receive accurate content even when there is no human in the loop. “In some cases when the answer is needed right away, customers can use our machine translations based on a model that still gives extremely excellent outputs,” Orqueda said.

In just a few years, LILT has grown its business and developed models that deliver a quality product for customers. “Looking ahead, LILT will continue delivering an extremely good experience for customers by using the best possible models and also creating the right training data for their particular use cases,” concluded Orqueda. “The next stage for us is making LILT a completely open platform where every single company can offer their multilingual services.”