Problem
AI assistants can propose and execute useful work, but opaque actions make them difficult to trust in operational contexts.
currently building · Product / Frontend Prototype
Relay AI
An agentic operations interface that exposes plans, tool calls, approvals, execution states and recovery—not another opaque chatbot.
Product decisions
Relay separates planning, tool use and execution, requiring explicit human approval for consequential actions.
UI system
Plan, streaming progress, tools, evidence, permissions, errors and recovery are first-class interface states.
Engineering
The prototype targets a Next.js client, server-side model gateway, structured outputs and typed tool contracts.
Challenges
The system must communicate uncertainty and partial failure without overwhelming users with internal detail.
Solution
Progressive states and review checkpoints keep people oriented, informed and in control.
Result
An active learning project, not a claim of professional AI production experience.
Next iteration
Implement one bounded end-to-end workflow with streaming, real tool calls, trace evaluation and recovery tests.
Interested in the thinking behind this work?
Start a conversation