Prototyping Generative AI Solutions at the University of British Columbia Cloud Innovation Centre

The University of British Columbia (UBC) Cloud Innovation Centre (CIC) has emerged as a vibrant incubator for creativity and technology, particularly focused on generative AI applications. Through collaborative initiatives with public sector sponsors, UBC CIC has utilized AWS (Amazon Web Services) to tackle real-world challenges via rapid prototyping methods. This article explores how the Center leverages Amazon Bedrock, shares valuable insights, and details the tools and methodologies that can aid your own AI projects.

The Significance of Rapid Prototyping in AI Development

Rapid prototyping is indispensable in the realm of artificial intelligence, especially for applications rooted in large language models (LLMs). Iteration is key; AI solutions require frequent refinements to enhance prompts, streamline data pipelines, and optimize outcomes. The UBC CIC has exemplified this with over 55 solutions developed, accentuating the importance of timely feedback from sponsors. Using Amazon Bedrock, students have created more than seven generative AI solutions, effectively saving time on infrastructure setup and maintenance.

Amazon Bedrock: Fueling Innovation

Amazon Bedrock offers a managed service that provides an array of models from leading AI companies and open-source communities all through a unified API. This platform equips the UBC CIC to build generative AI solutions while ensuring robust security, privacy, and responsible AI practices.

Ensuring Security in Generative AI Prototyping

Among the paramount concerns in AI development is data security. The UBC CIC adheres closely to the AWS shared responsibility model, making use of Amazon Bedrock’s comprehensive data protection features. Notably, Amazon Bedrock does not store or utilize customer data for training AWS models, preserving user privacy in accordance with UBC’s principles.

Access Control and Safety Measures

The UBC CIC employs AWS Identity and Access Management (IAM) for stringent access control, ensuring that only authorized personnel can interact with specific models or resources. This capability allows UBC CIC to prevent unauthorized access, enhance project-specific model usage, and ensure better management of provisioned resources during development. Regular security updates from AWS further bolster the integrity of generative AI applications.

To manage content safety, the Center uses Amazon Bedrock Guardrails, which effectively filter inappropriate topics, block harmful content, and eliminate personal identifiable information (PII) from both inputs and outputs. This safeguarding is critical in developing secure applications that must navigate sensitive data landscapes.

Selecting Optimal AWS Services for Generative AI Prototyping

The UBC CIC’s success is rooted in the strategic selection of AWS services, allowing students to focus on solving problems rather than juggling infrastructure concerns. Some key services they utilize include:

Amazon Bedrock

With straightforward testing and deployment capabilities, Amazon Bedrock empowers UBC CIC prototypes. The access to a diverse range of LLMs, such as Meta’s Llama and Mistral AI, allows students to dedicate their efforts to addressing sponsor-driven challenges.

Serverless Architecture

Adopting a serverless architecture reveals several advantages, including:

• Cost Efficiency: Sponsors pay only for the resources actually used.
• Scalability: Services like Amazon Bedrock and AWS Lambda adapt according to demand.
• Rapid Development: The UBC CIC can follow two-week development cycles, allowing for swift iterations and progress updates.

Utilizing LangChain

UBC CIC employs LangChain for its generative AI applications, streamlining LLM interactions through an abstraction layer. This not only expedites development and testing but also simplifies updates as model versions advance.

Evaluating LLMs for Specific Use Cases

The approach at UBC CIC involves careful evaluation of LLMs, targeting problem domains and identifying the smallest effective models. They rigorously test pre-trained models hosted on Amazon Bedrock, guided by clear project goals. A/B testing, using predefined prompts, allows for direct performance comparisons of LLMs.

Students also utilize Amazon SageMaker notebooks to refine prompts, seamlessly transitioning validated prompts into the prototype phase with the help of AWS Lambda for processing and Amazon API Gateway for interaction management. This creates an efficient pathway from testing to deployment.

RAG and Vector Storage Solutions

In terms of data management, UBC CIC harnesses pgvector with Amazon RDS for PostgreSQL. This powerful combination facilitates fast similarity searches across text embeddings, aligning cost-effective vector storage with relational database capabilities for enhanced performance.

Environment Setup Simplified

The UBC CIC leverages Infrastructure as Code (IaC) via the AWS Cloud Development Kit (CDK). By coding infrastructure in TypeScript, students can create and modify their setups rapidly, ensuring consistency and version control. This agility allows for prototype deployments in a fraction of the time—averaging only two hours instead of days.

Exemplifying Generative AI Projects

One notable collaboration is with the UBC Faculty of Science Advising Office, where they developed the Student Advising Assistant. Initially relying on an LLM hosted on Amazon Elastic Compute Cloud (EC2), they later migrated to Amazon Bedrock, substantially enhancing response quality while reducing costs by over 50%.

Additionally, in partnership with the UBC Centre for Teaching, Learning and Technology, they launched the AI Learning Assistant (AILA). This tool encourages proactive student engagement with course material while providing insights into comprehension levels. The serverless nature of AILA, powered by Amazon Bedrock, ensures that it accommodates comprehensive academic support effectively.

To conclude, the UBC CIC stands as a model for what’s achievable in generative AI innovation, fueled by AWS and supported by dedicated research and development. From security protocols to effective prototyping techniques, their framework not only fosters rapid advancement in AI solutions but also ensures that security and privacy uphold the integrity of their work. For further insights and guidance on building secure and scalable generative AI applications, visit the UBC CIC portal and explore Amazon Bedrock.