High-Level Design: Demand Planning for Subscription Grocery (Low-Capex)

Designing a demand planning and inventory management system for a subscription grocery retailer requires balancing the extreme volatility of perishable goods against the constraints of modest capital expenditure (Capex).

Based on recent advancements in fresh food supply chain research, this architecture eschews expensive physical automation (like massive AS/RS systems) in favor of cloud-native predictive intelligence, low-cost edge sensors, and human-centric AI.

1. System Architecture Overview

The system operates on a cloud-native, microservices architecture designed to minimize upfront hardware costs. It integrates three core modules:

Probabilistic Forecasting Engine: Predicts subscription base demand and dynamic add-on behaviors.
Dynamic Inventory & Routing: Manages FEFO (First-Expired-First-Out) and daily replenishment.
Cold Chain Telemetry: Tracks environmental conditions using low-cost edge networks.

graph TD
    A[Subscriber App / Web] -->|Orders & Subscriptions| B(Order Management System)
    C[Weather & Exogenous Data] --> D{Probabilistic Forecasting Engine}
    B --> D

    D -->|Prediction Intervals| E[Inventory Optimization Module]

    F[BLE Pallet Sensors] -->|Temperature Data| G[IoT Edge Gateway]
    G -->|Statistical Process Control| E

    E -->|FEFO & Markdowns| H(Warehouse Operations)
    E -->|Time-Dependent Routing| I(Last-Mile Delivery)

    style D fill:#f9f,stroke:#333,stroke-width:2px
    style E fill:#bbf,stroke:#333,stroke-width:2px

2. Probabilistic Forecasting Engine

Traditional point-forecasting fails for perishables because the cost of overstocking (waste + carbon penalty) is asymmetric to the cost of understocking (lost sales).

For a low-capex environment, massive deep learning models (like Temporal Fusion Transformers) are overkill and expensive to maintain. Instead, research indicates that tree-based machine learning models are highly effective and resource-efficient for short-horizon grocery data.

[!TIP] Algorithm Selection: Utilize XGBoost or LightGBM trained with a Pinball Loss Function. This directly outputs probabilistic prediction intervals (e.g., "There is a 90% chance we will need between 400 and 450 units of leafy greens"), allowing inventory decisions to be quantitatively linked to risk tolerance.

Feature Engineering Inputs:

Base Subscription Data: Recurring orders establish a strong baseline.
Exogenous Regressors: Weather forecasts, local events, and day-of-week seasonality to predict "add-on" purchases (e.g., a spike in berries on sunny weekends).

3. Inventory & Order Optimization

Inventory theory for perishables dictates that we cannot rely on standard FIFO (First-In-First-Out).

The Shift to FEFO

Implementing FEFO (First-Expired-First-Out) is structurally superior for reducing waste. By strictly rotating stock based on calculated expiration rather than receiving date, the system minimizes shrinkage.

[!IMPORTANT] The Cost of Overstocking: The algorithmic cost of overstocking ( $c_o$ ) must reflect not just the purchase price minus salvage value, but also the explicitly tracked disposal costs and carbon penalties associated with modern ESG reporting.

Dynamic Markdowns

For items nearing the end of their shelf life, the system uses dynamic markdown heuristics. Because this is a subscription model, these near-expiry items can be automatically surfaced to customers in the app as deeply discounted "Imperfect" or "Zero-Waste" add-ons just prior to their delivery cutoff.

4. Low-Capex Cold Chain Telemetry

Maintaining the cold chain is critical for shelf-life, but outfitting an entire fleet and warehouse with expensive active RFID or continuous cellular loggers violates the low-capex constraint.

The Solution:

BLE (Bluetooth Low Energy): Deploy low-cost BLE sensors at the pallet or tote level. These sensors act as digital Time-Temperature Integrators (TTIs).
Edge Gateways: Use fixed LoRaWAN or WiFi gateways at warehouse chokepoints (receiving docks, loading bays).
Anomaly Detection: Instead of complex digital twins, use Machine Learning (like Isolation Forests) for statistical process control on the incoming data streams to flag temperature abuse without raising false alarms.

5. Human-Centric AI (Supply Chain 5.0)

A common pitfall is attempting to build "Supply Chain 4.0" systems—fully autonomous, "black box" algorithms that automatically place orders without human oversight. These systems are incredibly expensive to build safely and often fail when dealing with unprecedented disruptions.

Supply Chain 5.0 embraces human-centricity. The AI serves as an advanced decision-support tool.

[!NOTE] Augmentation, Not Replacement: The system should automatically handle 90% of routine replenishment using the probabilistic engine. However, the system's primary UI should focus on surfacing the 10% of high-variance exceptions (e.g., a sudden heatwave or supplier shortage) to human demand planners, presenting them with clear, quantified options rather than making unilateral decisions.