Crypto.com Delivers Accurate Sentiment Analysis in 1 Second with Generative AI on AWS

Crypto.com was founded in 2016 with a bold mission: to bring cryptocurrency to every wallet. The company distinguishes itself from other trading platforms through its strong focus on driving user adoption, with a wide network of partners including merchants and payment gateways. Crypto.com currently serves around 100 million users in 90 countries.

With such rapid growth and a diverse customer base—all within an evolving, competitive industry—Crypto.com has embraced generative artificial intelligence (AI) to quickly deliver optimized customer experiences. The company’s AI use cases include conversational assistants for onboarding and customer queries, plus wizards to generate social media marketing campaigns.

In generative AI applications, sentiment analysis and news narrative categorization are becoming increasingly vital; particularly in the cryptocurrency space, investors need timely and accurate market intelligence to make informed decisions. Crypto.com provides a market insight service where users can access the latest news and information from both crypto and traditional news sources to gain insights from the company’s intelligence engines. Each subscription is tailored to the user’s trading level and the coins in their wallet.

Crypto.com employs a range of off-the-shelf machine learning (ML) models alongside its own custom models for sentiment analysis. However, developers encountered challenges with the limitations of open-source models and the high cost of self-hosting large language models (LLMs). They also faced accuracy issues with outputs generated by open-source models, especially when dealing with multilingual news sites. Consequently, Crypto.com sought a more effective solution to integrate and synthesize outputs from multiple ML models—both pre-trained and custom—to provide accurate, reliable, and comprehensive insights into the crypto market.

Crypto.com has run on Amazon Web Services (AWS) since its launch, so when it sought new LLMs for sentiment analysis, Anthropic Claude 3 on Amazon Bedrock was a logical choice. These LLMs are highly scalable and process vast amounts of data in real-time, facilitating comprehensive market research. Initial results showed that LLMs on Amazon Bedrock returned results very quickly, typically within one second.

Integrating Amazon Bedrock eliminated the manual effort, extra cost, and computational constraints of self-hosting LLMs. Within one month, Crypto.com had implemented Anthropic Claude 3 Haiku models on Amazon Bedrock for sentiment analysis, collecting and analyzing crypto news in more than 25 languages. The firm also utilizes Amazon Bedrock for ongoing proofs of concept (POCs) and development. Sunny Fok, head of AI & innovation technology at Crypto.com, says, “Generative AI on AWS services like Amazon SageMaker and Amazon Bedrock streamlined our adoption of the latest LLMs and AI technologies. We can now take innovative ideas from POC to full-scale production in weeks.”

To ensure domain-specific knowledge is applied in model output, Crypto.com used its own data to fine-tune open-source models including Mistral AI and Meta Llama on Amazon Elastic Compute Cloud (Amazon EC2). This approach is crucial when a new coin appears on the market, as off-the-shelf models often deliver subpar results. Crypto.com then began using Amazon SageMaker as an on-demand, end-to-end ML development platform to fine-tune its custom models.

Raymond Lam, senior engineer at Crypto.com, says, “Amazon SageMaker provides the tools and APIs to easily adapt our models with a user-friendly interface. Similar to Amazon Bedrock, we only need to run machine learning jobs as needed, and it’s much easier to manage custom modeling than if we were doing it on our own.”

Crypto.com received extensive support from AWS throughout the idea generation, POC, and production phases of multi-agent deployment. "We had several discussions on what tools or frameworks to deploy for different use cases,” says Lam. “The AWS team shared existing use cases, provided sample code, and demonstrated each step. When technical issues came up, AWS solutions architects helped troubleshoot and offered suggestions, which sped up our onboarding to generative AI on AWS services."