Subway – AI-Enabled Process Optimisation
Applied machine learning in a live retail environment to optimize operational processes. Delivered measurable improvements in efficiency and customer experience through intelligent automation and data-driven insights.
Applied ML
Real-world machine learning deployment in high-traffic retail environment
Live Deployment
Active system processing thousands of customer interactions daily
Process Optimization
Intelligent workflow automation reducing food waste and operational costs
Project Overview
The Challenge
Subway faced significant operational inefficiencies across multiple locations and dynamic food traffic flow: inconsistent of demand over different locations and times of the day, paired with events in the surrounding area, made a challenge about staffing, food orders and overall scheduling of the consumables in the shops.
Traditionally regional managers have to manually calculate requirements for staff workers and food orders based on previous weeks and knowledge of events in different areas. The processes was time consuming and prompt to mistakes given the complexity of the requirements.
The Solution
I developed an Machine Learning system based on YOLO (You only look once), which tracked different food items consumption during events and regular business hours. Paired with data from venues near, Subways' own data for sales, the system to analyzes customer flow patterns, predicts demand, and provides real-time recommendations to staff for preparation timing and resource allocation.
In a dashboard the system gives regional managers an overview of predicted consumption in euro value, type of sandwiches and ingredient needed. On separate side there is actual consumption in comparisor of the predicted so they have an overview when the system deviate and with how much.
Technical Implementation
Computer Vision Pipeline
YOLO-based detection for customer counting, queue length, and item consumption by sandwich type.
Demand Forecasting
Hierarchical time-series models (1h, 4h, daily) combining in-store signals with events and historical sales.
Closed-loop Learning
Feedback-driven optimization adjusts recommendations based on deviation between predicted and actuals.
Real-time Dashboard
Predicted vs. actuals in € value, sandwich mix, and ingredient needs for regional planning.
POS Integration
Connectors for existing POS and inventory systems with Excel import, validation, and schema mapping.
Edge Processing
In-store compute for low-latency inference and privacy-preserving, anonymized aggregation.
Results & Impact
for regional managers when it comes to staff rotation and orders planning
Through smarter distribution and ordering process
Optimized staffing and scheduling
Improved service consistency