The Challenge
The client, a leading national Third-Party Logistics (3PL) provider, faced escalating last-mile delivery costs, significantly impacting their profit margins. With an extensive network spanning urban and rural areas, delivering on time while minimising cost was becoming increasingly complex. Their existing manual route planning processes were proving inefficient, prone to human error and failed to adapt to rapidly changing traffic conditions, delivery window constraints, and dynamic customer requirements. The client needed an advanced technology-driven solution to maintain competitive advantage and meet growing customer expectations.
Our Solution
Adyantrix stepped in to address these challenges by developing an AI-based Route Optimisation Engine specifically tailored for logistics operations. Our solution leveraged cutting-edge data analytics and machine learning algorithms to automate and optimise route planning, ensuring the most efficient paths for delivery fleets. We integrated real-time traffic data, weather forecasts, and vehicle capacity into the system, allowing dynamic route adjustments to cater to immediate shifts in the logistics landscape.
Our Route Optimisation Engine also featured robust cloud-based infrastructure to ensure scalability and fast processing across the client's vast geographical coverage. We incorporated a user-friendly interface for ease of adoption and enabled API integrations with their existing Transportation Management System (TMS) for seamless operations.
Key Features
- AI-Driven Dynamic Routing: Utilises machine learning to continuously enhance route planning based on a multitude of variables, including historical delivery data and current road conditions.
- Real-Time Adjustments: Integration of real-time data sources to instantly adjust routes as new information becomes available.
- Scalability: Cloud-based architecture capable of handling a large volume of deliveries across varied geographical locations without compromising performance.
- Improved Fleet Visibility: Enhanced tracking and reporting features to provide logistics managers with comprehensive insights into delivery performance and areas for improvement.
- User-Friendly Interface: Intuitive design ensures quick adoption and ease of use for logistics operators.
Results
The implementation of the Route Optimisation Engine led to remarkable improvements for the client. Most notably, the last-mile delivery costs were reduced by 28%, significantly boosting their profit margins. The optimisation brought a 40% increase in the number of deliveries made within promised time windows, leading to improved customer satisfaction and repeat business.
Moreover, the automated routing processes eliminated manual planning inefficiencies, saving critical time and reducing operational errors. The seamless integration of our system with the client's TMS facilitated smooth communications across their logistics ecosystem, increasing operational transparency and coordination.
Ultimately, Adyantrix's solution provided the client with a competitive edge by enhancing delivery accuracy and reliability while significantly cutting down costs. This robust, scalable, and intelligent routing engine has equipped them to handle present and future logistics challenges with confidence and efficiency.
Technical Approach
The Route Optimisation Engine was built as a Python-based microservices application deployed on AWS, using a combination of ECS Fargate for containerised workloads and Lambda functions for event-triggered route recalculation tasks. The core optimisation algorithm was an implementation of a genetic algorithm extended with local search heuristics—specifically a 2-opt improvement pass applied after each generation—to solve the Vehicle Routing Problem with Time Windows (VRPTW) across the client's daily delivery manifest, which typically contained between 1,800 and 3,200 stops across a mixed urban-rural geography.
Real-time traffic data was sourced through the HERE Traffic API, with road speed data refreshed every five minutes and ingested into a Redis cache layer to ensure the routing solver always worked from current network conditions rather than static historical averages. Weather data from the Met Office DataPoint API was used to apply dynamic speed reduction factors on rural routes during adverse conditions, reducing the incidence of missed delivery windows caused by weather-related delays that the previous manual system had failed to anticipate.
The machine learning component was a gradient-boosted regression model trained on 22 months of historical delivery data, predicting stop-level dwell time—the time a driver spends at each delivery location—based on features including postcode sector, delivery type, parcel count, time of day, and day of week. Accurate dwell time predictions are the single most important input to VRPTW solving quality: underestimating dwell time leads to routes that look feasible on paper but consistently miss delivery windows in practice. Our model reduced dwell time prediction error from a mean absolute error of 4.2 minutes per stop (using the client's previous fixed lookup table) to 1.7 minutes, a 60% improvement that had a direct and measurable impact on route feasibility accuracy.
Implementation Highlights
The project ran over six months and was structured in four phases. The first phase focused on data preparation: extracting and cleaning 22 months of historical delivery records from the client's legacy TMS, which stored route and stop data in a normalised relational schema that had accumulated significant data quality debt over several years of operation. Approximately 31% of historical stop records required geocoding correction before they could be used as training data, as the legacy system stored customer addresses as free-text strings rather than validated coordinates.
The second phase was model training and solver development. The genetic algorithm was benchmarked against three alternative approaches—simulated annealing, Clarke-Wright savings, and Google OR-Tools' built-in VRPTW solver—across a test set of 60 representative daily manifests drawn from the cleaned historical data. The genetic algorithm with 2-opt local search consistently produced solutions within 2.3% of the OR-Tools benchmark on solution quality whilst running 40% faster, making it the preferred choice for a production environment where routes needed to be recalculated in under 90 seconds to meet the client's operational planning window.
Phase three was a parallel-running validation period of eight weeks, during which the engine generated optimised routes daily but the client's planning team retained the option to override them. Override rates started at 34% in week one and declined to 9% by week eight as planners built confidence in the system's recommendations. The remaining 9% of overrides were analysed and used to introduce additional business rules—primarily relating to specific customer access restrictions and driver knowledge requirements for certain rural delivery areas—that the algorithm had not been able to infer from historical data alone.
Phase four was the full live deployment, integrated directly with the client's TMS via a REST API, with routes published to drivers through the existing mobile scanning application. No new driver-facing hardware or applications were required, which significantly reduced the change management burden and allowed the full fleet to adopt the new routing capability within a single operational day.
Measurable Outcomes
The 28% reduction in last-mile delivery costs broke down across three contributing factors. Fuel consumption per delivery dropped by 17% as a direct result of shorter, more efficient routes with fewer redundant vehicle miles. Vehicle utilisation improved as routes were balanced more evenly across the fleet, reducing the number of vehicles required to clear the same daily manifest by an average of 11 vehicles per day—a saving that deferred a planned fleet expansion the client had been evaluating. The third factor was a reduction in penalty charges from customers whose service level agreements specified financial penalties for late deliveries: the 40% improvement in on-time delivery performance reduced these charges by approximately 62% year-on-year.
Driver feedback, collected through post-route surveys administered via the mobile scanning application, indicated that 78% of drivers found the new routes to be more logical and less stressful than the manually planned routes they had previously followed. Fewer instances of drivers needing to backtrack or make unscheduled judgment calls about stop sequencing were reported, contributing to a 12% reduction in driver-reported end-of-shift fatigue scores across the first quarter of live operation.
The system's real-time adjustment capability also had a measurable impact on exception management. When a vehicle breakdown or traffic incident affected an active route, the engine automatically recalculated the affected driver's remaining stops and redistributed undeliverable consignments to adjacent drivers with available capacity, reducing the number of failed delivery attempts requiring re-presentation the following day by 44%.
Lessons Learned
The most valuable lesson from this project was that data quality is the primary determinant of optimisation quality—not the sophistication of the algorithm. The eight weeks invested in geocoding correction and historical data cleaning during phase one paid dividends throughout the project. Every improvement in training data quality produced measurable improvements in model accuracy, and the planning team's confidence in the system's recommendations was directly correlated with the system's ability to generate routes that matched their operational experience of the delivery network.
The parallel-running validation phase was also essential, both technically and organisationally. Technically, it surfaced business rules and local operational knowledge that could not be inferred from data alone. Organisationally, it gave the planning team agency in the adoption process: they were not required to trust a system they had not been involved in validating. The decline in override rates from 34% to 9% over eight weeks was not a concession—it was evidence that the system was earning trust incrementally, which produced a more robust and sustainable adoption outcome than a hard cutover would have delivered.
Why This Approach Worked
The Route Optimisation Engine succeeded because it was designed as an operational tool rather than an analytical report. Previous optimisation tools evaluated by the client had produced high-quality route recommendations on historical data in batch mode but were too slow to generate routes within the client's live planning window and too rigid to accommodate the real-time exceptions that are a constant feature of last-mile logistics operations.
By combining a fast, heuristic-enhanced solver with a real-time data ingestion layer and an API integration that published routes directly into the existing driver workflow, Adyantrix built a system that was both technically capable and operationally usable from day one. The decision to deploy on AWS with auto-scaling configured on the ECS cluster also proved its worth during the pre-Christmas peak period, when daily manifests increased by 140% above the annual average: the solver scaled horizontally to maintain route generation times within the operational planning window without any manual infrastructure intervention.
Speak with our Custom Software Development team at Adyantrix to find out how we can support your next project.
Work with Adyantrix
If you are looking to tackle a similar challenge, Adyantrix has the expertise to help across the full project lifecycle. Our custom software development practice covers tailored applications built to your exact workflows. Our data analytics practice covers BI reporting and self-serve analytics platforms. Our AI & machine learning practice covers ML model development, MLOps, and intelligent automation. Our cloud & DevOps practice covers cloud infrastructure, CI/CD, and platform engineering. Get in touch to discuss your requirements — no commitment required.
