Smart Supply Chain Network Optimisation (Datathon 2025)

Project Overview

In a 24-hour datathon, we converted messy GPS shipments into a tractable hub-and-corridor network and ran a prescriptive, multi-objective optimisation (Gurobi MIP) that trades off Cost, On-time, and CO₂. The output is a clear decision story with three executable policies, corridor flow maps, and scheduling lists that managers can act on immediately.

Methods

• Network construction: KMeans clustering yields 12 hubs; each hub connects to its 3 nearest neighbours to form realistic, sparse corridors. This compresses raw points into a computable network while preserving geography.
• Arc coefficients: for every corridor we estimate unit shipping cost (c_ij), expected delay hours (d_ij), a CO₂ proxy from fuel consumption (e_ij = fuel × 3.16 kg CO₂e/L), and capacities — all stored in an optimisation-ready table.
• Prescriptive optimisation: minimise λ₁·Cost + λ₂·Delay + λ₃·CO₂ under supply/demand, flow conservation, capacity, SLA/delay caps, and emission caps. Traverse the frontier either by tuning weights (λ’s) or by imposing hard constraints.
• Policy simulation: switch weights/caps to realise Cost-first, On-time-first, and Low-carbon-first regimes. Under tighter SLA/CO₂ caps, cost rises and some instances become infeasible — the slope of that curve is the marginal cost of stricter targets.

What I Did

• Owned analytics & modelling end-to-end: baseline diagnostics, segmentation, KPIs, hub–corridor design, and MIP formulation/solve.
• Built decision artefacts: policy comparison tables, corridor flow maps, three routing schedules, and an exec-ready storyline.
• Coordinated with time-series teammate to align ARIMA/ETS+reconciliation attempts with network KPIs and final exhibits.

Reflection

• Balanced network: small deltas (±~1%) across policies indicate a reasonably balanced corridor system — optimisation still clarifies day-to-day choices.
• From prediction to prescription: GBM/ARIMA (R²≈0) struggled in noisy conditions; prescriptive optimisation turns signals into actions.
• Frontier intuition: tuning λ’s or adding hard caps makes the cost of SLA/ESG explicit — a teachable curve for stakeholders.
• Data → trust: estimating c_ij/d_ij/e_ij from first principles improved explainability and reviewer confidence.
• Stress tests: tightening delay/CO₂ caps surfaced feasibility boundaries early, valuable for scenario planning.
• Story matters: clear maps/tables/schedules + a crisp narrative moved the decision faster than algorithms alone.