Platform
One engine. Four problems. Many surfaces.
Quincus is a single optimization and simulation core. The capabilities below are problem classes it solves, not products it ships. Sector platforms are thin surfaces over the same engine.
Platform architecture
Six layers. One decision.
Every layer above estimation consumes distributions rather than point values. That single choice is why the platform degrades gracefully when the network misbehaves, and it is the reason the optimization layer can carry a bound rather than a guess.
How it fits together
Thin surfaces. Shared intelligence.
Each sector surface is a commercial wrapper, deliberately thin and independently replaceable. The engine, the data, and the patent sit in one place and compound across every surface above them.
Learned priors
The model proposes. The solver disposes.
Quincus trains network representation models on eleven years of shipment trajectories. They do not replace the optimizer. They make it faster and start it closer to the answer, while the optimizer keeps the guarantees a learned policy cannot provide.
Representation
A graph encoder over the network produces embeddings for facilities, lanes, carriers, and commodities.
Policy
An attention decoder produces a construction policy over route actions, trained on historical and simulated instances.
Guarantee
The learned output enters as a warm start and a branching prior. Feasibility and bounds continue to come from the exact layer.
A learned policy alone does not certify optimality and does not guarantee feasibility under hard time windows. We use learning where it compounds, which is speed, warm starts, and demand and duration estimation, and we use exact methods where the guarantee is the product. Any vendor claiming a neural network alone solves constrained routing to optimality is describing a demonstration, not an operation.
Go deeper