An Advanced Generative AI Platform for Faster Research

Our client is a leading psychological research organization in the United States, representing over 157,000 researchers, educators, clinicians, and students. With an extensive knowledge base that spans numerous publications, articles, and psychological research papers, they needed to build a generative AI platform to streamline access to this information. The platform would enable researchers to easily search through a vast amount of resources across multiple languages while ensuring high accuracy and contextual relevance in the search results.

• Improving knowledge discovery for researchers across the client’s web platform with a cognitive search feature.
• Transitioning from ChatGPT 3.5 to Claude, optimizing the platform’s search experience to support multi-turn conversations and maintain context across discussions.
• Enabling the platform to handle multilingual inputs while maintaining the contextual accuracy of results.
• Implementing a Bias-Free and Inclusive Language Solution to address potential biases across different language models and ensure reliability.
• Saving time for researchers by simplifying the data search process across the client’s extensive database.
• Training new AI models on both existing and incoming data while integrating them into the client’s applications.
• Seamless deployment and scaling of AI models, while maintaining high-performance standards.

• Selected and trained LLMs: The platform’s search platform transitioned from ChatGPT 3.5 to Claude, allowing for optimized search experiences. Additionally, the platform uses 10+ language models, including those from Anthropic and Amazon Bedrock, supporting flexible model selection for diverse research needs.
• Integrated multilingual data handling: Advanced LLM models were used to support multilingual inputs, ensuring high contextual relevance while maintaining bias-free outputs.
• Embedded knowledge base: We also embedded the Client’s knowledge base into the platform, leveraging Amazon Titan Multimodal Embeddings to enhance the AI’s ability to generate contextually accurate responses.
• Deployed and scaled AI models: Claude and other LLMs were deployed using Amazon SageMaker and AWS Fargate for scalable, high-performance operations, supported by efficient container orchestration via Amazon ECS.
• Implemented performance monitoring: Amazon CloudWatch was integrated for real-time monitoring, enhancing visibility of platform performance.
• Set up vector database for continuous learning: Amazon OpenSearch was integrated as a vector database, enabling continuous learning from user feedback, refining the AI models over time.
• Automated integration and delivery: AWS CodePipeline and AWS CodeBuild were implemented to automate the system, with CI/CD enhancements facilitated through AWS Amplify.
• Deployed security measures: AWS Secrets Manager was used to ensure compliance with security protocols and protect the platform from online threats.

• 45% reduction in operational costs: Automation and AI integration streamlined processes, leading to a 45% reduction in operational costs, particularly in data management.
• 30% savings in resource allocation: The generative AI platform reduced the need for manual data searches, resulting in a 30% savings in resource allocation.
• 80% faster knowledge discovery: The platform significantly improved the speed at which researchers could find relevant information, improving overall research efficiency.
• 70% more accurate AI models: With precise training on psychological data and user feedback, the accuracy of the AI models improved dramatically.
• 35% enhanced contextual relevance: Claude’s advanced generative AI capabilities improved the contextual relevance of search queries by 35%.
• Bias-free language processing: The Bias-Free and Inclusive Language Solution ensured that the system’s outputs were consistent and free from biases.
• 65% improved security: With AWS Secrets Manager in place, the platform’s security and the integrity of sensitive research data improved by 65%, ensuring safe and reliable operations.

• Amazon Bedrock LLM Models: We utilized Bedrock’s few-shot learning capabilities to train generative AI models on the client’s vast database, supporting accurate understanding of complex research queries.
• Anthropic Claude: Replacing ChatGPT 3.5, Claude became the core of the platform’s cognitive search, optimizing performance and context retention in multi-turn conversations.
• Amazon S3: A reliable storage solution for housing the knowledge base, storing thousands of XML files on psychological research articles.
• Amazon SageMaker: Used for deploying and fine-tuning LLM models, including Claude and others, ensuring optimal performance and accuracy.
• AWS Fargate: Provided necessary scalability through a containerized platform to support high traffic and workload.
• Amazon ECS: Managed container orchestration for handling workload distribution and efficient resource management.
• Amazon CloudWatch: Enhanced performance monitoring and troubleshooting for the platform in real-time.
• Amazon OpenSearch: Used as a vector database to continuously improve AI model responses based on user feedback.
• AWS CodePipeline: Automated the build and deployment process, ensuring a smooth release cycle for AI model updates.
• AWS CodeBuild: Supported code compilation and testing, ensuring robust and reliable system performance.
• AWS Amplify: Provided CI/CD capabilities, enabling ongoing improvements and updates to the platform.
• AWS Secrets Manager: Protected the platform from cyber threats and ensured compliance with security protocols.