AI Friction Labs: Claude-powered Learning Simulation Platform on AWS

The platform began as a rapid prototype on a managed backend-as-a-service with a hosted AI gateway. We migrated the product to infrastructure owned and controlled by the client. The new AWS environment includes Amazon RDS for PostgreSQL, ECS Fargate with an Application Load Balancer, S3 and CloudFront for static delivery, ElastiCache for session state, and SES for transactional email.
This gave the client greater control over their data and infrastructure, while creating a scalable foundation for production workloads and future growth.

• Data migration: Migrated the existing application data to PostgreSQL without data loss.
• Infrastructure-as-Code: Built reusable Terraform modules and automated deployment pipelines with CodePipeline and CodeBuild, making infrastructure repeatable and deployments more consistent and auditable.
• Environment parity: Established consistent development, staging, and production environments to reduce deployment risk as the product evolved.

At the core of the platform is its Claude-powered conversation engine, a conversational AI for an education system built on Amazon Bedrock. Claude Sonnet and Haiku are used for different workloads, balancing reasoning capability, response speed, and cost efficiency. Each bot can be configured with its own system instructions, resistance rules, and knowledge base, allowing educators to create simulations for different subjects and scenarios without changing the underlying platform. This later allows for expanding from traditional education into professional training verticals (medical, legal, hospitality, sales) through configuration alone.

• Model orchestration: Claude Sonnet and Haiku deployed through Amazon Bedrock for production AI workloads.
• Knowledge grounding: Per-bot knowledge bases with relevance-based retrieval to keep responses aligned with the material provided by educators.
• Real-time interaction: Server-Sent Events (SSE) stream model responses to the browser for a responsive, live conversation experience.
• Safety fencing: Prompt-injection guardrails separate untrusted student input from the bot’s core instructions, helping preserve the intended behavior of each tutor even under active persuasion attempts.

Our team also built the workflows surrounding the AI conversation, turning an individual chat into a complete learning and assessment experience. Students can set goals and review case materials before a session, interact with the bot during the simulation, and reflect on their performance afterward. Educators can then review the conversation and add annotations directly to individual messages.

• Pre-brief: Structured goal setting and scenario materials, including PDF case dossiers.
• Debrief and self-assessment: Post-session reflection and self-scoring, followed by a controlled score reveal with mandatory learner agreement capture.
• Educator review tooling: Inline annotations anchored to individual messages for targeted feedback and review.
• Voice interaction: Turn-based speech input and spoken playback using Amazon Transcribe and Polly, expanding the platform beyond text-based simulations.
• Document handling: Secure PDF and PPTX uploads with text extraction, so submitted materials can be incorporated into the AI’s context.