Back to architecture
live

Vertical AI Agent

A ReAct loop that quotes, samples, and hands off to a human.

The agent runs a reasoning loop: it thinks, picks a grounded tool (catalog lookup, quote, sample request), and acts, repeating until it can help. When a visitor is ready, it proposes a quote or sample and hands off to the owner over Slack and email rather than transacting on its own.

Data flow

How the pieces connect and where data moves.

Components: Visitor (states a need), Agent loop (ReAct reasoning), Gemini (plans next step), Tools (quote, sample, catalog), Catalog + docs (grounded facts), Owner handoff (Slack + Resend). Connections: Visitor to Agent loop; Agent loop to Gemini for reason; Gemini to Agent loop for next action; Agent loop to Tools for call tool; Tools to Catalog + docs for look up; Catalog + docs to Tools for facts; Tools to Agent loop for result; Agent loop to Owner handoff for quote / sample; Agent loop to Visitor for reply.

Loading data-flow map

Behavior

How it behaves step by step over time.

Loading behavior diagram
The agent loops: it asks Gemini for the next step, runs a grounded tool such as a catalog lookup, quote, or sample request, and feeds the result back. When the visitor is ready it hands off the quote or sample to the owner over Slack and email, then confirms with the visitor.

Operating guardrails

Constraints enforced in code, described at category level.

  • SDS, hazard, and handling answers come only from tool-returned documents, never model-generated.
  • Prices, products, and quotes are grounded in retrieved catalog data, never invented by the model.
  • Human-in-the-loop: the agent proposes quotes and samples and hands off to the owner; it never autonomously transacts or commits pricing.
  • Tenant isolation: every query is company-scoped, with no cross-tenant data.
  • Data privacy: visitor memory is self-deletable, messages are retained at most one year, and summarized memory is injection-defended.
  • Cost governance and anti-abuse: per-tenant token metering, caps, and rate limiting on the agent surface.
  • Transparent qualification tracks known and unknown facts; models are Google Gemini tiers.