FleetHappens

The Starting Point

It began with a blank Markdown file — MY_PROJECT.md — and a question: what if fleet data could tell stories?

Geotab gives you trips, GPS breadcrumbs, and raw telemetry. What it doesn't give you is meaning. A driver stops at a location 14 times in 90 days, and no one notices. A fleet manager wants to brief a driver about an unfamiliar area, and there's nothing to reach for. A tourism operator wants to share a route with passengers, and the best they have is a spreadsheet.

That observation became the kernel. The idea: FleetHappens — a full-stack web app that reads Geotab data and layers on context, intelligence, and narrative. Think of it as giving a fleet a voice.

The Vibe Coding Approach

This wasn't built line by line. It was built with AI, using a deliberate multi-agent architecture inside Cursor.

The process started with a Technical Writer + Solution Architect agent that read every competition resource and sharpened the project spec. That spec became the single source of truth. From there, a Tech Lead / Orchestrator Agent broke the work into parallel tracks and assigned them to specialist agents:

• Frontend / UX Agent — Next.js pages, Tailwind layouts, component shells
• Geotab Direct API Agent — auth, devices, trips, GPS breadcrumbs, device status
• Ace Integration Agent — the async create-chat → send-prompt → poll loop
• Context Briefing Agent — geocoding, Places API, LLM narration
• Story Generation Agent — LLM comic generation, tone system, PDF export

Each agent had its own prompt, its own file ownership boundaries, and a clear set of deliverables. The orchestrator kept them coherent. I acted as the human in the loop — tuning prompts, reviewing outputs, and enforcing quality.

The Build Timeline

Stage 1: Scaffolded the entire project from scratch. Vehicle selector, trip map with GPS breadcrumbs, Geotab Direct API integration, basic dashboard. Got to a working “select a vehicle, see its trips on a map” in a single day.

Stage 2: Upgraded story generation, integrated Ace with proper polling and retry logic, set up BigQuery for caching Ace results and storing comic stories. Added per-user Geotab authentication. By end of day the app was running on Google Cloud Run.

Stage 3: Added Fleet Pulse dashboard, vehicle intelligence cards, next-stop prediction, location dossiers, Street View, text-to-speech narration, PDF export, and a natural-language AI assistant. Zero to production in a matter of days.

What Vibe Coding Actually Felt Like

There's a version of vibe coding that's just autocomplete. This wasn't that.

The meaningful parts were the decisions that couldn't be delegated:

• Recognizing that Ace's 30–90 second latency would kill the stop-click experience — and designing a two-phase briefing that makes Phase 1 feel instant
• Figuring out that Geotab uses x for longitude and y for latitude, and centralizing that normalization before it could infect every component
• Deciding that the LLM is never the source of truth for coordinates, distances, or fleet data — it only generates narrative from structured inputs
• Choosing withFallback() as a universal wrapper so the demo could survive a network outage at the presentation

The agents wrote the code. I wrote the architecture, the constraints, the prompt engineering, and the integration rules. The interesting part of the journey wasn't the typing — it was the thinking.

What I'd Tell Someone Starting Now

1. Start with a spec, not a scaffold. The time spent on MY_PROJECT.md before writing a line of code paid back tenfold.
2. Design for demo reliability first. Every API route wrapped with a fallback. The live demo can fail gracefully without anyone noticing.
3. Treat the AI like a junior engineer, not a magic box. The agents needed boundaries: file ownership, interface contracts, explicit rules about which APIs handle which data.
4. The “wow moment” is worth protecting. The stop-click → context briefing flow was the single interaction worth fighting for. When scope creep threatened the timeline, asking “does this protect the wow moment?” was the right filter.
5. Vibe coding doesn't mean low standards. It means high velocity. The AI handles the boilerplate so you can spend your limited time on the parts that actually matter.