Child-safe multi-agent chatbot (elliotx.games)

A Minecraft-themed site for my son with a chatbot that routes questions to specialist agents, built with safety guardrails for a young user.

The problem

Generative AI for a child has different requirements from generative AI for an adult: the system has to be safe by design, not by hoping the model behaves, and the experience has to make sense to someone young.

What I built

Multi-agent routing: the chatbot, Bloki, routes each question to a specialist agent based on what is being asked. The handover is shown to the user as a familiar “transferring you to an agent” moment rather than hidden.
Guardrails: safety and security are architectural requirements. Each agent has constrained behaviour appropriate to a young user.
Infrastructure: a static site on AWS Amplify, with the chatbot backend on AWS Lambda behind API Gateway and the domain managed through Route 53. The backend moved from Netlify functions after hitting rate limits.
Testing: a Playwright and axe-core suite covers accessibility.

What I learned

Showing the routing to the user, instead of hiding it, made the product easier to understand and easier to test.
Serverless was the right fit for a chatbot with bursty, low traffic, and moving providers was cheap because the function boundary was already clean.