CASE 03 // 2025
Fleet System
Enterprise fleet intelligence at scale.
TrackFleet AI is an enterprise fleet platform I led at Techcora: live geolocation, fuel and over-speed alerting, geofencing, and predictive maintenance, scaled to 800+ vehicles on a Rust backend before the project was discontinued.
“Fleet operators went from zero visibility to live telemetry on 800+ vehicles, with predictive maintenance flagging failures before they happened.”
— Gagan Raj
CONSTRAINTS
- Telemetry from hundreds of vehicles arriving concurrently, around the clock
- Operators needed alerts in seconds, not on the next page refresh
- Fuel theft and route deviation are adversarial: the system had to be hard to fool
- Small team, so the stack had to stay operable by few hands
THE BUILD
The client ran a fleet in the hundreds of vehicles with no live picture of it. Fuel went missing, routes drifted, and maintenance was whatever broke first. I led the build of TrackFleet AI end-to-end: ingest, alerting, dashboards, and the prediction layer.
I had discovered Rust two months into the project year and made the call to put the ingest and alerting path on it: Axum and Tokio handling the constant telemetry stream into PostgreSQL, with the operator-facing app in Tauri and React. The alerting engine watched fuel curves, speed, and geofences in real time; the maintenance model flagged vehicles trending toward failure so they came in on schedule instead of on a tow truck.
The platform scaled past 800 vehicles in production before the project was discontinued for commercial reasons on the client side. The system held until the last day. I keep it on this page because the numbers were real, the architecture did its job, and shipping something that works is a separate skill from choosing what survives commercially. Both are worth showing.
DECISIONS — INCLUDING THE REJECTED ONES
Ingest path
CHOSE Rust with Axum and Tokio
REJECTED Node.js, which the team already knew
Telemetry ingest is exactly the workload where Rust pays rent: predictable latency under constant concurrent load, no GC pauses in the alert path.
Alert evaluation
CHOSE Streaming evaluation as points arrive
REJECTED Cron-style batch scans of recent data
A fuel-theft alert that arrives 15 minutes late is a report, not an alert. Operators act on seconds.
Desktop app
CHOSE Tauri wrapping the React control room
REJECTED Electron
Same web UI, a fraction of the memory, and the Rust side of Tauri shares types with the backend.
OUTCOME
- Scaled to 800+ vehicles in production
- Live geolocation with fuel, over-speed, and geofence alerting
- Predictive maintenance flagged failures before breakdown
- Led end-to-end until the project was discontinued
CRAFT LEDGER
- Axum/Tokio ingest
- Streaming alert engine
- Geofence evaluation
- Predictive maintenance model
- Tauri control room