How We Added ETA Prediction and Delay Alerts to an Existing Logistics Tracking App
Last Updated on October 13, 2026
Quick Summary
This insight explains how we retrofitted dynamic ETA prediction and automated delay alerts onto a logistics tracking app that was already running in production. It covers why the original ETA was wrong, how the prediction model was built from data the app already collected, how delay alerts were designed to stay useful rather than ignored, and how the features were released without interrupting live dispatch.
Table of Contents
Introduction
A live tracking map answers one question: where is the shipment right now? The question customers and dispatchers actually ask is different: when will it get here, and when that answer changes, when will they be told?
Our client ran a mid-sized regional fleet with a logistics tracking app their drivers and dispatchers already used every day. The app did its core job reliably. What it could not do was tell anyone that a delivery was going to be late until it already was. This insight covers what we changed, in the order we changed it.
Why the Existing Logistics Tracking App Couldn't Predict Delays
The platform was not broken. It was built for visibility, and visibility is a different problem hfrom prediction.
The ETA shown in the app was calculated once, at dispatch, from planned route distance and an average speed assumption. It never moved after that. A truck could sit in a yard for 90 minutes and the app would still display the original arrival time.
Four specific gaps came out of the discovery phase:
Position without context: GPS pings arrived every five minutes and were stored as coordinates. Nothing in the data model distinguished a truck stopped at a customer dock from one stopped in traffic.
No record of dwell time: Arrival and departure at each stop were never derived, so the app had no history of how long a given site typically took to load or unload.
Delays were discovered by phone: Dispatchers learned about problems when a driver called or a customer complained, which was usually after the promised window had closed.
No concept of a commitment: The app stored a planned arrival time but had no field for what had actually been promised to the customer, so nothing could be measured against it.
The last point mattered most. Without a committed window stored against each stop, there was no definition of late for the system to act on.
How We Implemented Dynamic ETA Prediction
We treated the ETA as something that gets recalculated continuously rather than issued once. The work ran in four parts.
Step 1: Turning Raw Pings Into Events
Before any modelling, we built a processing layer that converts the GPS stream into trip events: departed origin, en route, stopped, arrived at stop, dwelling, departed stop. Geofences around each customer site made arrival and departure detectable, which produced the dwell-time history the platform had never had.
Step 2: A Travel-Time Baseline From Live Road Conditions
For the road segment between the truck’s current position and the next stop, we call a routing service that returns duration under current and historical traffic. Google’s Routes API exposes selectable traffic models, including a pessimistic option, which we use for shipments with tight delivery windows where an optimistic estimate is the more costly error.
Step 3: A Correction Model Trained On The Fleet's Own History
Road-network travel time does not account for this fleet’s realities: dock queues at specific customers, loading patterns by time of day, and route-specific behaviour. We trained a gradient boosting model on 14 months of historical trips to predict the gap between the routing estimate and the actual arrival. The final ETA is the routing baseline plus this learned correction. Our machine learning development team kept the feature set deliberately small, using only inputs the app already captured reliably: stop sequence, customer site, day and hour, vehicle type, historical dwell at that site, and remaining distance.
Step 4: Recalculation Triggers
Recomputing on every ping was wasteful and made the displayed ETA flicker. We recalculate when the truck departs a stop, when it has been stationary off-site for more than ten minutes, when live traffic on the remaining route shifts beyond a threshold, and otherwise every fifteen minutes.
The app shows a single time, but stores a confidence range alongside it. That range is what the alerting logic reads.
How We Built Automated Delay Detection and Real-Time Alerts
An alert is only useful if someone can still act on it. Our rule was simple: every alert has to arrive while a decision is still possible, and go to the person who can make it.
Defining a delay
A stop is flagged when the predicted arrival exceeds the committed window by more than a configured tolerance, which the client sets per customer. A 20-minute overrun matters for a scheduled dock appointment and does not for a flexible residential drop.
Severity tiers
Delays are classified by how much slack remains:
Watch: predicted arrival is inside the window but the margin has fallen below 25%. Visible on the dispatcher dashboard, no notification sent.
At risk: predicted overrun beyond tolerance, more than an hour before the window closes. Push notification to the dispatcher, who can still reroute or call ahead.
Confirmed late: overrun with under an hour of slack. Dispatcher and the customer’s notification channel both receive it.
Cause tagging
Each alert carries a probable cause derived from the event stream: traffic on the remaining route, extended dwell at the previous stop, a late start from origin, or an approaching driver break. Hours-of-service limits are part of this picture, since FMCSA rules require a 30-minute break after eight cumulative hours of driving and cap the daily driving window, which makes some delays entirely predictable hours ahead.
Suppression rules
The first version of the alert engine sent too much. We added deduplication per stop, a cooldown so an unchanged status cannot re-alert within 30 minutes, and an escalation-only rule that re-notifies only when severity increases. Alert volume dropped by roughly two-thirds, and dispatchers started reading them.
Need Your Tracking App to Predict Delays, Not Just Report Them? Hire mobile app developers to integrate AI-powered ETA prediction and delay alerts into your existing logistics platform, using shadow testing and feature flags instead of a rebuild.
Rolling Out the ETA Prediction Without Disrupting Live Operations
The app was carrying real deliveries throughout. Nothing shipped as a hard cutover.
Shadow mode first: For six weeks the model ran in production and wrote predictions to a table that no user could see. We compared every prediction against the actual arrival before anyone relied on it.
Accuracy gate before exposure: The new ETA stayed hidden until it beat the old static estimate on median absolute error across every region, not just overall.
Staged exposure: Dispatchers saw the predicted ETA first, with the original estimate still displayed beside it. Customer-facing notifications came later, after a full month of internal use.
Feature flags by region and account: Each capability could be switched off for one depot without a release, which meant a problem in one region never required a rollback for everyone.
A visible override: Dispatchers can pin a manual ETA. Every override is logged with a reason, and those logs became the most useful source of model improvements in the first quarter.
What Changed in the Logistics Tracking App After Launch
The table below compares how the platform behaved before and after the work.
Capability
Before
After
ETA calculation
Set once at dispatch, never updated
Recalculated on trip events and traffic changes
Dwell time
Not derived from the GPS stream
Captured per stop and used as a model input
Delay discovery
Driver phone calls and customer complaints
Automatic flag with probable cause and severity
Customer communication
Manual, after the window was missed
Triggered before the window closes
Commitment tracking
No committed window stored
Per-stop window with configurable tolerance
ETA accuracy
Based on the initial estimated travel time
Updated using live traffic, trip progress, and dwell time
Dispatchers now start the day with the flagged stops instead of calling drivers for status. The measurable gain was lead time: most delays are now surfaced while a reroute or a customer call is still worth making.
The decision we would repeat is shadow mode. The one we would change is building cause tagging later than the alerts themselves. An alert that says a shipment is late without saying why forces the dispatcher to call the driver, the exact loop the project set out to close.
Conclusion
Adding prediction to a working logistics tracking app is less about the model than about the data underneath it. The event layer, the dwell history, and a stored definition of what was promised had to exist before any ETA could be trusted, and those three pieces took longer to build than the model itself.
As a Logistics Application Development Company, Bacancy Technology approaches this kind of work as an extension of a live system rather than a rebuild. If your logistics tracking app already shows position and your team still finds out about delays by phone, the first step is usually a short audit of whether the tracking data you are already collecting can support prediction.
Enough to cover seasonal and weekly variation on your actual routes, which in practice means roughly a year. Below that, the model learns recent conditions well and handles unusual periods poorly. Teams with less history can still run a traffic-aware baseline and add the learned correction once the data accumulates.
Usually yes. The prediction service reads the existing GPS stream and writes an ETA back through an API, so the core platform stays in place. The work that cannot be skipped is the event layer that turns raw pings into arrivals, departures, and dwell periods.
Tie every alert to a decision someone can still make. Severity tiers based on remaining slack, deduplication, cooldown windows, escalation-only re-notification, and a probable cause on each message are what separate an alert queue people read from one they mute.
Accuracy is worth measuring as median absolute error against actual arrival, segmented by route and customer, rather than as a single headline percentage. The practical benchmark is whether the prediction is stable enough and early enough for a dispatcher to act, which matters more than a few minutes of precision.
Rajiv Mehta
AI/ML Developer at Bacancy
Versatile tech leader driving innovation across cloud, AI, and full-stack development.