Trinity: Agentic AI-powered transition planning for students with disabilities
Trinity is a conversational AI solution built by University Startups that helps students with disabilities create personalized, IDEA-aligned transition plans through natural dialogue rather than static forms The system migrated from a single LLM prototype to a hierarchical six-agent architecture on Amazon Bedrock (Claude 3.5 Sonnet) to address hallucination risks, compliance requirements, and scalability challenges Key technical challenges included multi-domain context competition in a single pr
Analysis
TL;DR
- Trinity is a conversational AI solution built by University Startups that helps students with disabilities create personalized, IDEA-aligned transition plans through natural dialogue rather than static forms
- The system migrated from a single LLM prototype to a hierarchical six-agent architecture on Amazon Bedrock (Claude 3.5 Sonnet) to address hallucination risks, compliance requirements, and scalability challenges
- Key technical challenges included multi-domain context competition in a single prompt, lack of off-the-shelf retrieval tools for disability-specific college/career filtering, and strict HIPAA/FERPA compliance needs
- The serverless AWS architecture uses Amazon API Gateway, Lambda, DynamoDB, Cognito, Polly, Transcribe, and Bedrock Guardrails, with Canvas LTI 1.3 federation for school single sign-on
- Trinity reached over a dozen U.S. states in its first year and is expanding internationally into Saudi Arabia and Kuwait
Why It Matters
This case study demonstrates a practical, production-grade implementation of agentic AI in a regulated, accessibility-critical domain—offering a blueprint for how startups can responsibly scale AI systems that handle sensitive student data. It also highlights the importance of multi-agent architectures over monolithic LLM approaches when dealing with complex, multi-domain workflows requiring both accuracy and compliance.
Technical Details
- Multi-agent architecture: A hierarchical six-agent system with an Orchestrator Agent that sequences specialized agents (College Agent with 30,000+ record knowledge base, Employment Agent with 1,000+ occupation records, Training Agent) under a single coordinator to prevent cross-domain contamination
- AWS serverless stack: Amazon Bedrock (Claude 3.5 Sonnet) for AI inference, Amazon API Gateway (REST + WebSocket) for client communication, AWS Lambda for compute, Amazon DynamoDB for conversation state, Amazon Cognito with Canvas LTI 1.3 for authentication
- Accessibility layer: Amazon Polly for text-to-speech output and Amazon Transcribe for speech-to-text input, enabling full voice-based interaction without backend modifications
- Responsible AI & compliance: Amazon Bedrock Guardrails for content filtering and hallucination reduction, field-level encryption, strict role-based access controls, and automated data retention policies to meet HIPAA and FERPA requirements
- Universal Design for Learning (UDL): Product design specifications aligned with UDL standards to ensure the system is accessible to students with a wide range of abilities
Industry Insight
- Multi-agent architectures should be considered the default for production AI systems handling regulated or multi-domain workflows, as monolithic LLM approaches risk cross-contamination of context and increased hallucination rates
- Accessibility-first AI design (voice input/output, UDL compliance) is not just an ethical imperative but a technical differentiator that expands market reach, especially in education and government sectors
- The partnership model between mission-driven startups (University Startups) and specialized AI engineering firms (g/d/n/a on AWS) represents a scalable template for rapidly deploying compliant, production-grade agentic AI systems without building full internal ML engineering teams
Disclaimer: The above content is generated by AI and is for reference only.