Enterprise2021· EMSPL — India

EMSPL Fleet Intelligence

AI Fleet Management for 2,400 Vehicles

24-month enterprise engagement delivering an AI-powered fleet management platform for 2,400 vehicles with predictive maintenance and real-time GPS intelligence.

Headline outcome

2,400 vehicles · 12% fuel savings · 41% fewer breakdowns · 24-month partnership

A multi-level road interchange at night, traffic drawn out into light trails — Illustrative photography. No imagery on this site is from a client environment.

React NativeFastAPIGPS TelemetryPredictive MLOffline-firstRedis

Overview

EMSPL operates one of India's largest commercial fleets. They needed a modern platform to replace their legacy GPS tracking system with AI-powered predictive maintenance, driver behaviour analytics, and fuel optimisation.

Challenge

The fleet operated across 18 Indian states with inconsistent connectivity. The platform needed to work offline-first on vehicle-mounted tablets, sync intelligently when connected, and process real-time telemetry from 2,400 GPS units simultaneously.

Solution

We built an offline-first React Native app for drivers with background sync. A FastAPI backend processed 2,400 concurrent telemetry streams. AI models predicted maintenance requirements 14 days in advance based on vibration, fuel, and mileage data. Real-time dashboards for fleet managers.

Outcome

Fuel savings of 12% through AI route optimisation. Maintenance costs reduced 23% through predictive scheduling. Breakdown incidents dropped 41% in Year 1. 24-month partnership with ongoing feature development.

Measured

2,400

vehicles managed

41%

fewer breakdowns

12%

fuel saving

24 months

partnership, still running

The constraint that shaped it

The fleet does not stop for a release. Vehicles keep moving and keep transmitting during a deployment, and a driver on a route with no coverage still has to complete a log. That single fact made offline-first the default rather than a fallback, made ingest a queue that absorbs a backlog instead of a request that can fail, and made every prediction accountable in public — because a flagged vehicle that never failed costs a workshop day the depot will remember.

How it was built

How the numbers were measured

41% fewer breakdowns and 12% fuel saving are operational figures, so the comparison has to be an honest one: the same 2,400 vehicles, measured against the fleet’s own prior records rather than against a modelled projection. Predictive maintenance is then judged on two failures, not one. A missed fault is obvious, because the vehicle stops. A false flag is the quieter cost — a workshop day spent on a vehicle that was fine — and it is the figure that decides whether a depot manager keeps acting on the alert.

What the driver and the fleet manager each see

Two users, two different products. The driver has an in-vehicle application that has to work with no signal, so capture is local and sync runs in the background — offline is the normal case on a long route, not a degraded mode. The fleet manager gets live position and a maintenance queue. The decision that mattered was making a prediction arrive as a scheduled job with a vehicle, a reason and a date, rather than as a chart. Nobody books a service off a trend line. They book it off a work order.

Telemetry that never stops arriving

The hardest part was ingest, not prediction. 2,400 units transmit whether the platform is mid-deployment or not, and a truck that has been out of coverage for hours returns a backlog of fixes with a device clock that disagrees with the server. So a position is stored with both its device time and its receipt time, repeated fixes are idempotent on arrival, and a backlog is absorbed at the rate the database can take rather than at the rate it turns up. Deploy behind that and no vehicle loses a day of history.

From fleet platform to enterprise resource platform

The engagement did not end at handover; it changed shape. The fleet platform is extending into a full enterprise resource platform, with the ERP in beta from August 2026, delivered on a standing monthly retainer with joint IP. Twenty-four months in, it is still running. What that model removes is the re-scoping negotiation every time the roadmap moves. The same engineers already hold the data model, so a new module is priced in person-days against a codebase they know, not estimated against a description of one.

Stack, by concern

Telemetry & ingest

GPS telemetryStreaming ingest

Prediction

Predictive MLMaintenance schedulingFuel and route analysis

Applications

React NativeOffline-first syncFastAPI

Engagement

Dedicated team retainerJoint IPSLA-backed operations

More work

Krida AI

National Sports Management Platform

GamesTheShop

A custom commerce platform for India's largest gaming retailer

All case studies Start a project