Interactive demo · v0.1

Ask your fleet anything.

A small LLM agent that turns natural-language fleet questions into structured tool calls against a Turo dataset. The demo dataset happens to have five vehicles, but the schema and tool layer work for any fleet size. Bring your own Anthropic key for the real model, or play in demo mode with deterministic responses.

Why this work

The dashboard answered the questions I knew to ask. The agent answers the ones I haven't thought of yet.

A scorecard surfaces the metrics you preselected. A natural-language layer lets an operator (or a future me at 2 a.m.) ask "which car is bleeding cash this month and why" without knowing the schema. The hard part is not the model; it's the tool design, the eval loop, and keeping the LLM on rails. This page is the spec, the live demo, and the thinking behind both.

Craft on display
  • PRD writing
  • Agent architecture
  • Tool design (typed inputs / outputs)
  • Eval framework
  • Human-in-the-loop guardrails
  • LLM cost modeling
  • BYO-key + demo-mode UX
  • Spec craft for engineering handoff
Demo mode · no key required
claude haiku 4.5 · 8 tools
Fleet Agent
0 messages
Agentgreeting
Hey, I'm a small agent that can answer questions about a Turo fleet. The demo has five sample vehicles, but everything I do (schema-aware SQL, typed tools, structured output) works the same on a one-car fleet or a fifty-car fleet. Try a prompt on the right, or ask anything in the box below. In demo mode I follow scripted call paths so you can see the tool architecture without a key.
Related work
Built artifact
Vantage Fleet OS / interactive dashboard
The opinionated scorecard the agent calls under the hood. Same dataset, different surface area.
View dashboard →
Case study
Why I built Fleet OS
Why I rejected the spreadsheet, scoped the MVP, and what I would change next.
Read case study →