Data Science Project

Predictive Warehouse Optimisation

PULPO WMS — Predictive Warehouse Optimisation Engine

This project is shared with the written permission of PULPO WMS. Underlying client data, production metrics, and commercially sensitive results remain confidential under NDA. The dashboard shown runs on anonymised synthetic data generated for demonstration purposes.

About the Company

PULPO WMS

3PL & eCommerce Warehouse Management System

PULPO WMS is a cloud-native SaaS warehouse management system built for e-commerce merchants, 3PLs, and fulfilment operations. Operating across 20+ countries, PULPO provides an API-first platform covering picking, packing, inventory counting, returns, and incoming goods management.

This project was commissioned as a university collaboration to build a predictive intelligence layer on top of their existing platform.

20+

countries

Trusted by fulfilment experts

99.9%

uptime

Low-latency performance

2m 8s

response time

Support keeps warehouses moving

Modern architecture

Built on MACH principles

API-first microservices
Cloud-native SaaS
Headless design

Built for

To lead warehouse operations with a digital warehouse twin

D2C brands
3PL & fulfilers
Online merchants
Wholesalers

My Role

  • Project lead — coordinated across PULPO, the university, and a 4-person team.
  • Designed and built the product proximity model.
  • Integrated all 4 models into a single end-to-end pipeline.
  • Co-created EDA and helped with zone allocation model design.

The Opportunity

A predictive intelligence layer on top of the operational core

PULPO WMS already handles the operational layer — picking, packing, inventory. The project extended this with a predictive intelligence layer: using the rich transaction data PULPO captures to make smarter decisions about where products should live and when to reorder them.

Most warehouses organise products by category rather than demand velocity. By applying ML to PULPO's data, we could recommend zone placements that reduce picker travel time and automate replenishment before stockouts occur.

The Solution

A 4-model pipeline

Each model feeds the next — forecasts drive placement, placement drives replenishment — forming a single predictive loop on top of PULPO's operational data.

  1. Model 1

    Demand Forecast

    LightGBM

    Per-SKU, per-merchant demand forecast across 3 monthly horizons using LightGBM with a Tweedie objective — well suited to sparse, zero-inflated warehouse demand.

    OutputFeeds the zone allocation and replenishment models
  2. Model 2

    Product Proximity

    K-Means

    Clusters products by co-purchase behaviour using a sparse co-occurrence matrix enriched with category signals (70% co-occurrence, 30% category).

    OutputEnables the LP to co-locate frequently bought products in the same zone
  3. Model 3

    Zone Allocation

    Linear Programming

    PuLP/CBC solver maximises picks × zone efficiency subject to zone capacity constraints. Affinity weighting keeps co-purchase clusters together.

    OutputRecommended zone reassignment per product
  4. Model 4

    Auto-Replenishment

    Safety Stock + EOQ

    Calculates reorder points per product per zone using safety stock (Z=1.65, 95% service level), supplier lead time, and forecast demand.

    OutputTriggers classified by urgency — CRITICAL through LOW — with suggested order quantities

Live Dashboard

The pipeline, end-to-end

The dashboard below demonstrates the full pipeline end-to-end on anonymised synthetic data — production data is confidential.

Embedded AI Agent

Ask the pipeline anything

The dashboard includes an embedded AI agent powered by the DeepSeek API that answers natural language questions about any model output, warehouse metric, or configuration decision. Responses are grounded in the pipeline's actual outputs — bridging the gap between data science outputs and operational decision-making.

Why is a warehouse skipped?What does WAPE 0.56 mean?Which products should I prioritise?
⚠️ Metrics reflect synthetic data. Real client results are under NDA.

Key Takeaways

What the pipeline proved

Tweedie objective outperformed standard regression on sparse warehouse demand data.

Co-purchase clustering is more effective than category-based grouping for zone co-location.

LP zone allocation delivers meaningful throughput improvements even with conservative capacity constraints.