Ingress / Case Studies / Route Optimization, Demand Forecasting, Predictive Maintenance

Route optimization, demand forecasting, predictive maintenance.

A regional freight carrier reduced fuel costs by 18%, cut maintenance events by 34%, and forecast demand with 91% accuracy.

SectorLogistics & Freight
TypeAI / ML
Scale300 Employees
The Challenge

Operations run on spreadsheets and intuition.

A regional freight and logistics company with 300 employees and 180 vehicles was managing operations reactively: manual route planning, no demand forecasting, and fleet maintenance scheduled by calendar rather than condition.

The Situation

Dispatchers planned routes in Excel using historical knowledge and gut feel. No traffic prediction. No demand visibility beyond the next day or two. The 180-vehicle fleet was maintained on a fixed schedule, with no ability to detect failures before they happened. Result: excess fuel costs, unplanned downtime, and inefficient capacity allocation.

The company was growing, but operational costs were growing faster. Profitability was flat. Scaling required more vehicles and more fuel waste.

The Opportunity

Two years of shipment data, vehicle telematics, maintenance logs, and GPS tracking existed in separate systems. Combined and analyzed, that data could unlock three improvements: optimized routes that reduced distance and fuel, forecasted demand to match capacity ahead of time, and predictive alerts for maintenance before failures.

A machine learning approach could run continuously, adapting to weather, traffic, and seasonal patterns.

Our Approach

Three integrated ML pipelines.

We built three connected models, each feeding into dispatcher and operations dashboards.

01
Route Optimization
Python-based OR-Tools vehicle routing with live traffic APIs. Ingests current orders, vehicle locations, capacity constraints, and real-time traffic. Generates optimized routes that minimize distance and fuel. Dispatchers can accept or override. Integrated into existing dispatch software via REST API.
Weeks 1-8
02
Demand Forecasting
XGBoost model trained on 24 months of shipment history, augmented with external signals (weather, holidays, economic indicators). Predicts 7-day and 30-day demand by route, shipment type, and customer segment. Powers capacity planning. Refreshed weekly, monitored for drift.
Weeks 6-12
03
Predictive Maintenance
Random forest classifier on vehicle telematics (fuel efficiency, idle time, engine hours) combined with maintenance history. Flags vehicles likely to require service in the next 30 days. Operations team schedules proactive maintenance. Reduced emergency repairs and downtime.
Weeks 10-18
04
Dispatcher Dashboard
Streamlit dashboard displaying optimized routes, forecasted demand, fleet status, and maintenance alerts. Single pane of glass for operations. Updated hourly. Change management training over 2 weeks.
Weeks 16-20
The Outcomes

Measurable impact on every metric.

18%

Fuel & Route Savings

Optimized routes reduced miles per delivery by 18%. Annual fuel and operational cost savings: $1.7M (baseline $95M revenue).
34%

Fewer Maintenance Events

Predictive maintenance eliminated 34% of unplanned maintenance calls. Fleet availability up, downtime costs down.
91%

Forecast Accuracy

Demand forecast MAPE of 9% on 7-day horizon. Operations now confidently allocates capacity and optimizes staffing.
Tech Stack

Python, AWS, Streamlit, Airflow.

1

Python / OR-Tools

Google OR-Tools for combinatorial optimization. Route generation engine runs hourly on AWS Lambda, processing 200+ orders per cycle.
2

XGBoost / scikit-learn

Demand forecasting and predictive maintenance models. Trained on historical data, retrained weekly with drift detection (evidently.ai).
3

AWS (SageMaker, Lambda, S3)

Model hosting on SageMaker endpoints. Lambda for scheduled prediction jobs. S3 for data lake. CloudWatch monitoring and alerting.
4

Streamlit / Airflow

Dispatcher dashboard in Streamlit. Apache Airflow orchestrating daily data pipelines, model refreshes, and alerting.
Key Lessons

Building ML systems that stick.

Data Audit First

We spent 4 weeks understanding data quality, completeness, and alignment across 8 source systems. That upfront work uncovered missing telematics and incomplete maintenance logs, which we backfilled. Garbage in = garbage out.

Baseline Measurement Matters

Before deployment, we measured: average miles per delivery, fuel cost per mile, maintenance event frequency, forecasting accuracy of the manual process. These baselines made the impact quantifiable and visible to leadership.

Dispatcher Trust, Not Automation

Routes and maintenance alerts were recommendations, not mandates. Dispatchers could accept or override. Trust grew as the models proved accurate. Within 8 weeks, 92% adoption of suggestions.

Continuous Monitoring Essential

We built drift detection and model retraining into the pipeline. When forecast accuracy degraded (seasonal shift), Airflow automatically retrained and alerted ops. The system stayed accurate and trusted.

Start a conversation

Tell us what's worth doing.

// 30 minutes โ†’ a written brief.

Bring the problem. We'll come back with a written brief: what to build, what to defer, and where AI actually moves the number. No deck pitches.

Emailconnect@ingressits.com
GSA MAS#47QTCA26D000K
Reply< 24 hrs