March 18, 2025

Intelligence on the edge: Applying AI in the physical world of QSR

How Viam is helping quick service restaurants bring AI directly to physical locations, allowing businesses to monitor freshness and display fullness without straining networks or requiring expensive hardware upgrades.

.png)

The challenge: Pizza and AI in the real world

A quick service restaurant company selling pizza in malls and food courts faced a common challenge; their business model relies on displaying pizza slices for customers to view before ordering. Management identified two key opportunities to increase sales:

1. Ensure fresh product by replacing pizzas that had been displayed longer than 90 minutes
2. Keep display trays more than half full, as fuller displays were shown to drive higher sales

The company already had security cameras installed at their locations with views of the pizza displays. The question was how to leverage this existing hardware to provide analytics that could help increase sales without introducing significant new costs.

Why cloud-based solutions weren't the answer

The initial approach—streaming camera feeds to the cloud for vision processing—quickly ran into a critical constraint: network limitations. The restaurants used a single Wi-Fi network shared with their point-of-sale (POS) system. Past attempts at cloud-based vision overwhelmed the network, causing disruptive POS outages. Therefore, the retailer required any vision inference to happen on-site.

Moving inference on-site introduced another challenge: hardware cost. For a proof-of-concept, deploying expensive computing equipment wasn’t feasible. Any on-premises solution needed to run efficiently on low-cost hardware like Raspberry Pis.

Simpler approaches, such as periodic image uploads, were also impractical due to concerns about network reliability. Ultimately, the retailer required a vision solution that could operate independently without impacting their critical sales infrastructure.

Computer vision without cloud dependency

Viam's approach was to develop a specialized computer vision model capable of running directly on edge devices. Specifically, the model was:

• Lightweight enough for edge devices with limited computing power, such as Raspberry Pis
• Precise enough to reliably identify conditions like pizza freshness and tray fullness in real-time

By performing inference entirely at the edge, the solution removed the need for continuous cloud connectivity, ensuring reliability without straining network resources.

The key challenge then became: how could a small, specialized model be trained quickly enough to meet the retailer's needs without requiring extensive manual data preparation?

Model distillation: Extracting specialized intelligence

Rather than manually labeling data or building a model from scratch—a process that would typically require extensive manual data labeling—engineers at Viam accelerated development by using a vision-language model (VLM) to bootstrap the labeling process. Specifically, they:

• Collected representative images of pizza displays directly from existing security cameras.
• Used a VLM to automatically label these images, rapidly identifying pizza conditions and tray fullness.
• Trained a specialized, lightweight model on this auto-labeled dataset for efficient deployment on edge devices.

This approach enabled the team to leverage the generalized knowledge of a powerful cloud-based VLM and effectively distill it into a specialized, edge-deployable model.

Real-world results: AI at the edge

By deploying the specialized model directly on small computers at each location, the pizza retailer gained:

• Real-time analytics: Immediate insights about pizza freshness and display fullness without network latency
• Network independence: No impact on the critical POS system's network performance
• Automatic notifications: Alerts when pizzas needed to be replaced or displays needed to be restocked
• Low implementation cost: Leveraging existing cameras with minimal additional hardware investment

Most importantly, the solution worked within the practical constraints of a real-world quick service environment. Staff didn't need to learn complex new systems or workflows—they simply received actionable notifications about when to replace old pizza or restock displays.

Beyond pizza: The broader QSR opportunity

While this implementation focused specifically on pizza displays, the same edge-based AI approach can address numerous QSR challenges:

• Inventory management: Reducing the $1,000 in lost revenue per 3.3 pounds of food waste
• Service optimization: Preventing $1,000 in lost sales per day due to slow service or long lines
• Display management: Addressing the fact that 1 in 3 displays run empty during rush periods
• Equipment reliability: Mitigating the 40% of critical equipment failures that occur during peak hours

Specialized AI models running at the edge can bridge the gap between physical operations and digital intelligence without requiring constant cloud connectivity or extensive infrastructure changes.

The future of QSR operations

Bringing AI to physical operations doesn't require expensive infrastructure overhauls or constant cloud connectivity. With the right approach, existing hardware can be transformed into powerful data collection points, with lightweight, specialized AI processing happening directly where the business operates.

By connecting the physical world of restaurant operations to AI-powered insights, QSRs can make more informed decisions, reduce waste, improve service speed, and ultimately deliver better customer experiences.