Blueprintlogistics

From Static Planning to Dynamic Dispatch — AI-Powered Fleet Optimization for RD Energy

Q: How does this integrate with our existing TMS and ERP?

The platform connects via standard APIs (REST, OData, or file-based feeds for legacy systems) and acts as an intelligence layer on top of your existing stack — it does not replace your TMS or ERP. Integration typically takes 1–2 weeks per source system, with the data model normalized on our side so your IT team does not need to restructure anything.

Q: Do we need existing optimization models to use this platform?

No. IJONIS builds the complete stack end-to-end — from the MILP solvers to the AI orchestration layer. If you already run LP or network optimization models, we integrate with them. If you don’t, we build them as part of the engagement.

Q: How does the system handle situations where no feasible route plan exists?

When hard constraints (driver hours, vehicle capacity, delivery windows) make a plan infeasible, the system reports the infeasibility explicitly and identifies which constraints are in conflict. It then presents relaxation options ranked by impact — for example, extending a delivery window by 30 minutes or splitting an order across two vehicles — so the dispatcher makes an informed trade-off rather than receiving a silently degraded plan.

Q: What happens when the AI gets it wrong?

The AI layer classifies disruptions and interprets unstructured inputs — it does not generate route plans. Route plans are always produced by the deterministic solver. If the AI misclassifies a disruption severity or misinterprets an input, the worst outcome is that the solver is triggered unnecessarily or with slightly imprecise parameters. Every AI action is logged with confidence scores, and dispatchers can override any classification before the solver runs.

Q: How quickly does the demand forecasting model become accurate?

The model requires 8–12 weeks of historical delivery data to establish baseline consumption patterns per site. Accuracy improves continuously as actual sales data feeds back into the model. Within the first quarter, forecast-driven proactive replenishment typically reduces unplanned emergency deliveries by 30–40%.

Q: Can this be deployed on-premise?

Yes. The architecture supports cloud deployment, on-premise, or hybrid configurations. For organizations with strict data sovereignty requirements, the entire platform — including the AI inference layer — can run within your own infrastructure.

How an agentic orchestration layer bridges RD Energy’s existing LP models and real-time operations — combining deterministic optimization with AI-driven reoptimization across 800+ delivery points.

PythonGoogle OR-ToolsNext.jsSupabaseLangChainXGBoost

AI dispatch platform dashboard with fleet overview and route optimization

Project at a Glance

15–20%

Fleet cost reduction

80%

Less manual planning

800+

Delivery points

14 Days

Demand forecast horizon

<5 min

Reoptimization cycle

Project at a Glance

15–20%

Fleet cost reduction

80%

Less manual planning

800+

Delivery points

14 Days

Demand forecast horizon

<5 min

Reoptimization cycle

Case Study

Challenge

RD Energy distributes fuel from regional depots to over 800 retail service stations and commercial customers — mines, manufacturing plants, agricultural operations — across South Africa. Dispatchers build daily delivery plans for a mixed tanker fleet, balancing order urgency, vehicle capacity, driver regulations, site-specific delivery windows, and hazardous goods constraints.

The company uses linear programming models for network planning — optimizing depot allocation and transport costs across its distribution footprint. These models produce mathematically sound results, but they sit in the planning department, not in operations. By the time a plan reaches the dispatch floor, reality has already moved.

A mining customer calls in an unforecasted emergency top-up. A tanker breaks down on the N1 with 36,000 litres mid-route. A Gauteng service station sells through its forecast volume two days early during a cold snap. Each event invalidates the plan, and dispatchers respond with experience and phone calls rather than mathematics.

McKinsey estimates 15–25% of fleet operating costs in distribution networks are avoidable through dynamic optimization and demand sensing. What RD Energy is missing is not optimization capability — it’s an intelligence layer that connects planning to operations in real time.

Solution Architecture

The platform bridges RD Energy’s existing planning models and daily operations through four layers — each with a distinct role in turning static plans into continuously optimized execution. For organisations without existing optimization models, IJONIS builds the complete stack end-to-end — from the mathematical solvers to the AI orchestration layer.

Data Integration Layer

The platform connects to RD Energy’s existing systems — TMS, ERP, depot management, and vehicle telematics — through standardized APIs into a unified operational data model. Vehicle positions, order status, depot inventory levels, and delivery confirmations converge in real time. No manual exports, no spreadsheet handoffs. The integration layer also ingests external signals: traffic conditions, weather forecasts, and fuel price indices that affect delivery economics.

Demand Forecasting Engine

A machine learning layer analyses historical delivery patterns, seasonal trends, promotional calendars, and external variables (weather, public holidays, commodity price shifts) to forecast demand at the individual site level. For retail service stations, the model learns consumption curves per product grade — predicting when a site will approach stockout days before it happens, so replenishment can be scheduled proactively rather than reactively. Forecast accuracy improves continuously as the system ingests actual sales data, reducing the unplanned emergency orders that disrupt daily route plans at their source.

Optimization Engine (Deterministic)

At the core sits a mathematical optimization engine solving the Vehicle Routing Problem with Time Windows (VRPTW) — a well-established operations research formulation. The solver uses Mixed-Integer Linear Programming (MILP) to compute delivery plans that minimize total fleet cost under hard constraints: vehicle capacity, driver hours, hazmat routing rules, depot operating windows, and customer delivery slots. This is not AI — it is mathematics. The plans it produces are provably optimal within the defined constraints — the same class of solver RD Energy already trusts in its network planning models.

AI Reoptimization Layer (Agentic Orchestration)

This is where static becomes dynamic. An agentic AI layer monitors the live operational state against the current plan. When reality diverges, the system detects the deviation, classifies its severity, and decides how to respond:

Deviation type	Example	Response	Engine
Minor	15-min delay, small volume variance	Absorb — adjust timing within existing routes	Rule-based
Moderate	Vehicle breakdown, unplanned order	Reoptimize — re-solve affected routes via MILP with updated constraints	Optimization engine
Demand shift	Retail station trending toward early stockout	Pre-empt — trigger proactive replenishment run before stockout occurs	Forecasting + optimization
Major	Depot supply shortfall, multi-vehicle disruption	Orchestrate — simulate scenarios, recommend recovery plan, escalate for dispatcher approval	AI + Optimization + human-in-the-loop

The AI layer does not replace the optimization engine — it triggers it. The language model interprets unstructured inputs (a driver’s WhatsApp message about a road closure, a customer email changing a delivery window) and translates them into constraint updates that the solver can act on. The optimization engine then re-solves. Computed plans, AI-interpreted triggers — the boundary is always clear.

Core Capabilities

Real-Time Fleet Visibility

Unified view of vehicle positions, load status, and delivery progress across the entire fleet — replacing fragmented TMS and telematics screens

Dynamic Route Reoptimization

When disruptions occur, the MILP solver re-solves affected routes within minutes, factoring in updated constraints from the AI layer

Site-Level Demand Forecasting

ML-driven consumption predictions per product grade at individual retail and commercial sites — turning reactive stockout responses into proactive replenishment

Natural Language Dispatch

Dispatchers issue instructions in plain language — "prioritize the Mpumalanga mine deliveries today" — and the system translates intent into solver parameters

Scenario Simulation

Run what-if analyses in minutes: what happens if we lose a depot for 48 hours, add 50 new commercial customers, or cut the fleet by 10%?

Compliance-First Planning

Driver hours, hazmat routing, and site restrictions are hard constraints in every plan — the system cannot produce a non-compliant route

Security & Compliance

Fuel distribution is a regulated environment — hazardous goods transport, driver hours legislation, and site access restrictions are non-negotiable constraints. The platform treats compliance as architecture, not configuration.

Data Isolation

All data processing runs within the customer’s own cloud environment. RD Energy’s operational data — vehicle positions, order volumes, customer site information — never leaves their tenant. The AI inference layer runs on dedicated compute, not shared infrastructure. API credentials use short-lived session tokens invalidated after each sync cycle.

Regulatory Constraints as Hard Boundaries

Driver hours, hazmat routing exclusions, and site-specific delivery window restrictions are encoded as hard constraints in the MILP solver — they cannot be overridden by the AI layer or by dispatcher action. The system will not produce a route plan that violates these constraints, regardless of cost pressure. If no feasible solution exists within the regulatory boundaries, the system flags the infeasibility explicitly rather than relaxing constraints silently.

Audit Trail

Every plan generated, every reoptimization triggered, every dispatcher override is logged with timestamp, user identity, and the full constraint state at the time of the decision. When a regulator or internal auditor asks "why was this route chosen," the system produces the complete decision chain.

Learn more about the technologies and concepts behind this platform: AI Agent · Predictive Analytics · Data Pipeline · API Integration · Human-in-the-Loop

Projected Impact

15–20%

fleet cost reduction

80%

less manual planning

min reoptimization cycles

This is an engineered blueprint based on publicly available industry challenges. It does not represent work performed for any specific company.

Frequently Asked Questions

How does this integrate with our existing TMS and ERP?+

The platform connects via standard APIs (REST, OData, or file-based feeds for legacy systems) and acts as an intelligence layer on top of your existing stack — it does not replace your TMS or ERP. Integration typically takes 1–2 weeks per source system, with the data model normalized on our side so your IT team does not need to restructure anything.

Do we need existing optimization models to use this platform?+

No. IJONIS builds the complete stack end-to-end — from the MILP solvers to the AI orchestration layer. If you already run LP or network optimization models, we integrate with them. If you don’t, we build them as part of the engagement.

How does the system handle situations where no feasible route plan exists?+

When hard constraints (driver hours, vehicle capacity, delivery windows) make a plan infeasible, the system reports the infeasibility explicitly and identifies which constraints are in conflict. It then presents relaxation options ranked by impact — for example, extending a delivery window by 30 minutes or splitting an order across two vehicles — so the dispatcher makes an informed trade-off rather than receiving a silently degraded plan.

What happens when the AI gets it wrong?+

The AI layer classifies disruptions and interprets unstructured inputs — it does not generate route plans. Route plans are always produced by the deterministic solver. If the AI misclassifies a disruption severity or misinterprets an input, the worst outcome is that the solver is triggered unnecessarily or with slightly imprecise parameters. Every AI action is logged with confidence scores, and dispatchers can override any classification before the solver runs.

How quickly does the demand forecasting model become accurate?+

The model requires 8–12 weeks of historical delivery data to establish baseline consumption patterns per site. Accuracy improves continuously as actual sales data feeds back into the model. Within the first quarter, forecast-driven proactive replenishment typically reduces unplanned emergency deliveries by 30–40%.

Can this be deployed on-premise?+

Yes. The architecture supports cloud deployment, on-premise, or hybrid configurations. For organizations with strict data sovereignty requirements, the entire platform — including the AI inference layer — can run within your own infrastructure.