ZHOU AZ Live site ↗
Jarvis
A working AI finance agent surface: streaming tool calls, typed schemas, wallet-aware flows, structured outputs, and visible execution state.
What the system makes visible
Not a prompt trick. A product surface where users can see what the agent intends, which tools run, what comes back, and what remains under user control.
Five-minute script
001Start with the user jobA user wants an agent that can answer a finance question and move safely toward an executable workflow.
002Run a real queryAsk Jarvis to compare yield or token opportunities, explain assumptions, and reveal which tools it needs.
003Show tool executionPoint out tool names, typed inputs, intermediate statuses, structured responses, and error boundaries.
004Inspect the UIExplain why cards, tables, route previews, and transaction states are safer than burying tool output inside prose.
005Show extensibilityShow how a new tool enters the system, how state is handled, and where the interface needs to explain the result.
Architecture talking points
001RuntimeModel-driven flow through Vercel AI SDK with streaming chat UX.
002ToolsTyped definitions, custom registry, structured outputs, and per-tool UI result components.
003SafetyWallet/auth-aware surfaces, transaction previews, test behavior, and explicit confirmation boundaries.
004System craftConfiguration, integration boundaries, contribution conventions, and failure cases that can be inspected instead of guessed.