Trinity moves from one LLM to a six-agent system on Amazon Bedrock

University Startups has rebuilt Trinity, its conversational AI for postsecondary transition planning for students with disabilities, as a serverless multi-agent system on AWS. Instead of filling out static forms, students talk with Trinity about their interests, strengths and goals, and finish with a personalized transition plan aligned with the Individuals with Disabilities Education Act (IDEA). The product centers on the Individualized Education Program (IEP), a legally mandated document prepared by parents, teachers and specialists.

Founded in 2020, the organization works with middle and high school students across the United States. Trinity reached educators and students in more than a dozen U.S. states within its first year, and expansion into Saudi Arabia and Kuwait is underway. AWS partner g/d/n/a delivered the engineering build, from infrastructure and AI orchestration to compliance controls and the educator dashboard, while University Startups defined the product design specifications and Universal Design for Learning (UDL) standards.

Why the single prompt failed

Early prototypes put student intake, career exploration, compliance checking and plan writing into one prompt. In production, retrieval quality, instructional tone and regulatory requirements competed inside the same context window, so a failure in any one domain could corrupt an entire student session and produce hallucinations or inaccurate information. The team migrated to a hierarchical six-agent system, with specialized agents collaborating under a single orchestrator and a dedicated consolidation step assembling their outputs into one internally consistent plan.

The workflow separates recommendation from plan generation. In phase one, after a four-segment intake conversation, the orchestrator routes to the appropriate primary agent, which returns five to ten ranked options filtered to the student's context. In phase two, once selections are confirmed, every agent contributes its section of the plan.

The stack

A web client connects through Amazon API Gateway over REST and WebSocket, with compute on AWS Lambda. Amazon Bedrock with Claude 3.5 Sonnet powers the AI interaction and conversation state sits in Amazon DynamoDB. Amazon Polly reads responses aloud, Amazon Transcribe converts speech to text, and authentication runs through Amazon Cognito with Canvas LTI 1.3 federation for school single sign-on. Amazon Bedrock Guardrails handles content filtering and hallucination reduction.

Special education records must be protected under HIPAA and FERPA, so field-level encryption, strict role-based access and automated data retention controls were built in from day one. Full details are in the AWS blog post.