alirezaaminzadeh/omnitower-solver-config

Omnichannel Fulfillment — Solver Configuration

This repository publishes the solver configuration and benchmark summary for the Omnichannel Fulfillment and Returns Control Tower. It is a configuration / decision-model artifact, not a trained neural network: the "model" is a multi-stage optimization pipeline combining a greedy-marginal-cost order allocator, an Adaptive Large Neighborhood Search improvement loop, a bid-price-style dynamic admission controller, and a rule-based reverse-logistics disposition policy.

Components

Stage	Method
Inventory Allocation / Order Fulfillment	Greedy Marginal Cost heuristic over candidate sources (DC, regional WH, cross-dock, supplier, store)
Allocation Improvement	Adaptive Large Neighborhood Search (destroy random batch, reinsert by priority, accept if cost improves)
Assortment Substitution	Substitute-group lookup triggered on stockout
Dynamic Admission Control	Scarcity-scored accept / defer-to-backorder decision per incoming order
Multi-Option Routing	Fast / Economical / Green objective-weighted allocation for the same order
Reverse Logistics	Rule-based disposition: restock store / restock DC / repair / recycle / dispose

Cost Model

Term	Value (USD)
Per km, standard shipping	0.055
Per km, express shipping	0.145
Per unit picked	1.35
Per unit packed	0.85
Per split shipment	3.20

See config.json for the full manifest and the latest benchmark summary across all five demo network scenarios.

Intended Use

Decision-support and portfolio demonstration for omnichannel retail fulfillment and reverse-logistics planning. Not certified for live operational use; all data is synthetic.

Limitations

• Synthetic network topology and order stream, not a real retailer's data.
• No live carrier-rate, WMS, or OMS feed integration.
• Return quality scoring and disposition rules are simplified surrogates of a real grading/QA workflow.