CASE STUDY / LOGISTICS & SUPPLY CHAIN
Cutting delivery costs 14% across a national parcel network
Dynamic route optimization and ML-based ETAs across 3,800 vehicles — 14% lower cost per delivery and 96% on-time performance within three quarters.
14%
reduction in cost per delivery across the network
96.2%
on-time delivery, up from 90.8%
38 min
average ETA error reduced to 9 minutes
$21M
annualized operating savings verified by finance
THE CLIENT
Context
The client operates one of North America's largest middle- and last-mile delivery networks: 3,800 vehicles, 62 regional hubs, and 1.1 million weekly stops. Route planning was performed hub by hub, each morning, by dispatchers using a decade-old planning tool with static transit-time tables and heavy manual adjustment.
Volume growth from e-commerce contracts had pushed the network past what manual planning could absorb. Peak-season overtime was climbing 20% year over year, on-time performance had slipped below 91%, and the largest retail customer had introduced ETA-accuracy penalties into its renewal terms.
THE CHALLENGE
What was at stake
Three problems compounded each other. First, route plans left savings on the table: post-hoc analysis showed 12–18% excess miles versus optimal, driven by static assumptions about drive times, dwell, and stop density. Second, ETAs quoted to customers came from the same static tables and missed by 40+ minutes on a quarter of stops, generating where-is-my-delivery call volume and contract exposure.
Third — and hardest — any new system had to win over 240 dispatchers who had watched a previous optimization rollout fail. Plans that violated real-world constraints (dock schedules, driver route knowledge, customer quirks recorded nowhere digital) had destroyed trust in 'the algorithm' once already.
Engagement at a glance
- Client
- A Fortune 500 logistics operator
- Region
- North America
- Duration
- 9 months
- Team
- 11-person team: optimization engineers, ML, platform, and an embedded operations lead
Services applied
THE SOLUTION
What we built
We built a network optimization platform in three layers. The foundation is a real-time network state model: telematics pings, hub scan events, and order data streaming through Kafka into a geospatial store that always knows where every vehicle, package, and trailer is. This alone replaced a patchwork of nightly reports.
On top of it, a constraint-aware routing engine re-plans each hub's routes overnight and re-optimizes intraday when disruptions hit — encoding hours-of-service rules, vehicle capacities, customer time windows, and 130+ hub-specific constraints captured through dispatcher interviews. A gradient-boosted ETA model, trained on two years of telematics and delivery outcomes, replaced the static tables and publishes per-stop ETAs with confidence bands.
Critically, the system shipped as decision support: dispatchers reviewed recommended plans each morning, overrode freely, and every override was logged and reviewed weekly by a joint tuning team. Overrides fell from 31% of routes in month one to under 9% by month five as missing constraints were encoded.
// ARCHITECTURE
Event-driven core on Kafka with Flink stream processing for network state; PostGIS-backed geospatial services for location queries. The routing engine wraps OR-Tools with a custom constraint compiler so hub-specific rules are configuration, not code. ETA models are LightGBM ensembles served via a low-latency inference service, retrained weekly through an MLflow pipeline.
Everything runs on the client's AWS tenancy across three regions, deployed with Terraform and GitOps, with Grafana operational dashboards and full plan-versus-actual telemetry feeding the tuning loop.
Core stack
- Kafka
- Flink
- OR-Tools
- LightGBM
- PostGIS
- MLflow
- AWS EKS
- Terraform
- Grafana
- React
HOW IT WAS DELIVERED
Implementation approach
Value delivered in phases with go/no-go evidence at each gate — never a big-bang bet.
- 01
Diagnostic re-plan (weeks 1–6)
Re-planned eight weeks of historical routes for five hubs and compared against actuals — quantifying a 13.9% network-wide savings opportunity before any build commitment.
- 02
Data backbone (months 2–4)
Unified telematics, TMS, and scan events into the real-time network model; fixed location data quality at 14 hubs where geocoding errors would have poisoned optimization.
- 03
Two-hub pilot (months 4–6)
Ran optimized plans in production at two hubs with dispatcher override capture and weekly constraint-tuning sessions.
- 04
Network rollout (months 6–9)
Templated onboarding brought the remaining 60 hubs live in cohorts of eight, each with a two-week shadow period and local constraint capture.
“Previous vendors gave us a black box and told our dispatchers to trust it. Ilmora gave our dispatchers a better tool and earned the trust route by route. The savings are real — our CFO signed off on the number.”
Marcus Whitfield
SVP, Network Operations — Fortune 500 logistics operator
KEEP READING
More case studies
Want results like these?
Bring us the metric you need to move. Our architects will map how a comparable engagement would work in your environment — systems, timeline, and expected impact.