Geotab Geoff — AI Fleet Safety Coach with Lip-Synced Avatar (Vibe Coding Competition)

[robertlagrasse]

What I Built

Geotab Geoff is an AI-powered fleet safety coaching platform where a lip-synced avatar named Geoff has voice conversations with drivers about their actual shift data. Built for the Geotab Vibe Coding Competition 2026.

Live demo: geotab-geoff.web.app | Repo: github.com/robertlagrasse/geotab-geoff | Demo video: YouTube

How It Uses the Geotab APIs

Geoff uses all three Geotab data channels:

1. MyGeotab API (mg-api-js SDK v3.0.0)

• ExceptionEvent - fetches safety events (speeding, harsh braking, acceleration)
• Driver - maps events to drivers
• LogRecord - GPS coordinates for each event
• GetRoadMaxSpeeds - posted speed limits for context ("12 mph over in a 35 zone" vs just "speeding")

2. Ace AI (GetAceResults)

• Natural language query about the driver's history and patterns
• Key insight: Ace context is not one-shot. It gets injected into the Gemini coaching prompt and persists through the entire multi-turn conversation. So when a driver asks "what about last week?" Geoff already has that context.
• The async pattern: CreateAceChatAsync then GetAceResults then poll until complete

3. OData Data Connector

• Fleet-wide analytics for the supervisor dashboard
• KPIs, safety metrics, and trend data
• Feeds the fleetAnalytics Cloud Function

Architecture

Component	Tech
Frontend	React + Vite on Firebase Hosting
Backend	Cloud Functions v2 (Node.js 20)
AI Coach	Gemini 2.0 Flash via Vertex AI
Voice	Cloud TTS (Neural2-D) + Cloud STT
Lip Sync	Wav2Lip on Cloud Run with NVIDIA L4 GPU
Dashboard	MyGeotab Add-In (supervisor view inside MyGeotab)
MCP	6-tool FastMCP server for Claude Desktop

The Interesting Parts

Shift-level coaching, not event-level alerts. Geoff aggregates all events for a driver's shift and looks for patterns. GPS clustering via Haversine distance detects location patterns. Four speeding events at the same intersection means a signage problem, not a driver problem.

Three-tier escalation system. Data-driven (5+ events, 15+ mph over), conversation-driven (road rage, impairment, intentional violations), and driver-requested. Server-side safety net forces escalation when boolean flags fire, even if the LLM says do not escalate.

Cost: approximately $0.05 per coaching session. Gemini + TTS + GPU + STT + storage = about $1/driver/month for 100% coaching coverage. That is 500x cheaper than a human safety coach.

Built in approximately 48 hours with Claude Code. One developer, no team, no prior Geotab API experience. Full vibe coding journey here.