Focal Systems: Boosting Store Performance with an AI Retail Operating System and Real-Time Data

Every second a product sits out-of-stock on a shelf, revenue quietly drains away. Customers walk out empty-handed, and businesses lose customers as well as valuable insights into what’s actually happening on the floor. 

At Focal Systems, we have built Shelf AI, a system that continuously “sees” store shelves, detects stockouts, and guides teams to replenish in time—so products are on the shelf when customers need them.

Our mission is to modernize retail by automating store operations with computer vision, AI, and data streaming to maximize profitability and minimize food waste. AI isn’t an add‑on for us—it’s the core engine that powers our retail intelligence capabilities. By analyzing shelf images in real time, we help stores reduce waste and increase sales through accurate, immediate insights on shelf availability and store inventory. We achieve this by creating a simple, seamless solution for boosting store performance with an AI retail operating system that uses high-quality, real-time data at scale.

Every Image Provides Multiple Pieces of Data to Drive Decisions

In retail, freshness isn’t just for produce; it’s the real-time context that AI needs to act now. Our AI delivers trustworthy guidance if there’s context that gives meaning to the raw data—new shelf images, product and planogram changes, inventory and sales signals—and it flows continuously and is processed as it arrives. 

Data streaming is what lets us consume new shelf images, run model inference, compute aggregates, and publish insights and actions with minimal delay. That’s how we can point associates to the right aisle at the right moment instead of hours later. It’s also how managers get live shelf availability metrics for operational decisions that move the needle that same day.

The Focal System integrates cameras with deep learning models for real-time intelligence and consists of four main components:

1. Computer Vision: Proprietary discrete cameras designed specifically for the demands of the retail environment. Focal is able to cover the entire store, capturing images of every product on every shelf throughout the shopping day. 

2. Shelf AI: Focal’s proprietary AI engine interprets the data captured by our cameras to identify products that are in-stock versus out versus low so clients can see true product availability the same way their customers do. Our model spots mismatches between corporate planograms and what’s actually on store shelves (or not). See this video of Shelf AI cameras to learn more.

3. Action Tool: Focal’s Action Tool puts advanced AI at the fingertips of the store team, prioritizing and assigning tasks in real-time to improve compliance, product availability, and worker efficiency. Using the app, store workers are able to update inventory in real time, reducing food waste, and generating more accurate orders.

4. Impact: Impact is Focal’s management dashboard for measuring and managing the massive improvements computer vision can bring to retail operations: wasted labor redeployed, out of stock duration reduced, sales recouped, tasks completed, compliance enforced, and presentability improved. Impact includes Store Walk, Focal’s immersive platform that lets retail leadership virtually walk the aisles of any store from anywhere to gain a comprehensive view of product availability, shelf compliance, and operational performance.

In summary, the Focal Systems platform combines proprietary shelf‑mounted cameras, advanced AI, and real‑time data streaming to transform retail operations. Computer Vision cameras scan shelves continuously, Shelf AI interprets millions of images daily, the Action Tool mobile app guides associates to replenish stockouts, and the Impact dashboard provides live visibility and even time‑travel views of shelf conditions.

Challenges We Faced Building a Real-Time AI Retail Operating System

Before we standardized on a managed data streaming platform, we developed custom services and an internal framework that struggled to keep up with retail’s fast-paced demands, including:

• High‑cardinality data

  ‎ 

  It was difficult to process and visualize massive amounts of high-cardinality data–per store, per camera, per image–in an efficient way, in time to act.

  ‎ 

• Lack of scalability

  ‎ 

  Our custom framework had issues with handling high ingress data volume and cardinality, which stressed our point-to-point pipelines and made aggregate management inflexible–slowing down iteration on metrics and new features.

  ‎ 

• Needing enterprise-ready, managed solutions 

  ‎ 

  As a lean team, we needed a managed, highly available platform to accelerate development, but these solutions must also offer sufficient customizability to avoid becoming technical dead-ends as we scaled. Given our work with major enterprise retailers, enterprise readiness and guaranteed high uptime were non-negotiable requirements.

  ‎ 

• Balancing cost efficiency and unit economics

  ‎ 

  Cost efficiency is paramount; our architecture and technology choices are heavily evaluated against our internal unit economics model to ensure sustainable growth.

Using Confluent as the Streaming AI Backbone

A fully managed, cloud‑native, and scalable service was a hard requirement for us, and Confluent’s rich ecosystem accelerated our path from prototype to production without increasing operational burden. Because we were already using cloud-native Apache Kafka® on Confluent Cloud, that made it straightforward to enable serverless Apache Flink® alongside our existing stack.

• Event-driven streaming backbone: Kafka on Confluent is our communication layer across AI/ML pipeline stages and services in the image processing flow.

• Stream processing with Flink: We compute per‑store, per‑camera metrics in real time using Flink SQL and table abstractions mapped to Kafka topics.

• Connectors and governance: We rely on connectors for operational sinks (e.g., MongoDB) and for moving curated data into Snowflake; governance and cataloging are important as we expand datasets and teams interfacing with them.

Here’s the architecture at a glance:

At a high level, here’s how data moves from store shelf to actionable intelligence: 

1. Ingestion

   ‎ 

   Retail data lands via APIs into MySQL and streams into Kafka topics. This includes shelf images (and associated metadata), product updates, inventory levels, sales data, and more from our computer vision pipeline.

   ‎ 

2. Computer vision & inference

   ‎ 

   Our proprietary LLMs and ML models run on GCP. Image features and detections (e.g., product presence, gaps, misplaced items) are extracted and published to Kafka topics, then processed with Flink.

   ‎ 

3. Real‑time stream processing to make data AI-ready

   ‎ 

   While our main pipeline is image processing, we also process data from products, inventory, purchases, and model inference outputs. We use Flink to compute live metrics and aggregates—turning detections into shelf availability insights, out‑of‑stock events, and planogram compliance signals. The streaming layer is also our microservice communication fabric for the event‑driven image pipeline.

   ‎ 

   Pre‑AI data readiness steps include PII removal, masking, OCR, feature extraction, and other image preprocessing—ensuring that what reaches models and downstream systems is both privacy‑safe and model‑ready.

   ‎ 

4. Operational and analytical sinks

   ‎ 

   We write low‑latency operational results to MongoDB and stream curated datasets to Snowflake for analytics and reporting. We also use streaming for CDC to data warehouses and data lakes so that downstream systems stay current without batch ETL.

   ‎ 

5. Store Walk and Impact experiences

   ‎ 

   Actionable directives flow to our Store Walk mobile experience (guiding staff to replenish) and roll up as Impact metrics for store efficiency and shelf availability dashboards.

AI Use Cases for Retail Unpacked

Our AI stack includes multiple proprietary LLM and ML models. Choosing the right model for the job while keeping critical IP and training in‑house matters most.

Top applications powered by streaming data combined with our AI stack are:

• Shelf availability monitoring (out‑of‑stock detection)

  ‎ 

  Continuous inference turns image streams into reliable, time‑stamped signals that drive replenishment workflows in Action Tool.

  ‎ 

• Automated inventory auditing (cycle counting)

  ‎ 

  We compute discrepancies and trends from detections and metadata to reduce manual audits and improve accuracy.

  ‎ 

• Planogram compliance and merchandising optimization

  ‎ 

  We detect misplaced items and compliance issues off the live vision stream and surface the most impactful actions.

  ‎ 

• Store Walk “time travel”

  ‎ 

  Teams can review how any section looked throughout the day—paired with live computed metrics—because the data is processed and available in real time, not in tomorrow’s batch.

Business Impact of Real-Time Retail AI 

The biggest shift is that store decisions are driven by real-time context. New images arrive; we infer, aggregate, and publish—then associates act. Managers see the same grounded truth reflected in Impact, all in step with the store’s actual state. That virtuous loop depends on fresh, high‑quality context end‑to‑end, which is why streaming sits at the core of our architecture. Ultimately, Focal Systems is realizing the self-managing store of the future, driven by advanced AI.

The combination of managed Kafka and Flink on Confluent allows us to move quickly, simplify our streaming architecture, and focus on business logic rather than infrastructure management. We replaced custom services and previously unscalable frameworks with Flink SQL stream processing and accelerated delivery while lowering the barrier to production.

From a customer perspective, Focal Systems empowers store teams to receive timely, trustworthy directives through Action Tool, and retail stakeholders to get live shelf availability and inventory accuracy insights through Impact—both powered by enriched data streams that improve model outputs. The result is fewer stockouts, higher on‑shelf availability and inventory accuracy, faster replenishment and cycle counts, and increased sales—delivered with less waste and greater enterprise reliability across every store.

Get Started

Whether you’re building computer‑vision retail AI like we are—or any real‑time decisioning system that relies on fresh context—start by making the stream your system of record. Bring events into Kafka, progress continuously with Flink, and keep your operational and analytical stores fed by the same truth in motion. Then iterate quickly on the metrics that matter to your customers. 

Ready to build real‑time retail with us? Learn more about Focal Systems and how we’re transforming shelf availability and store execution with computer vision. Then get started on the data streaming foundation behind our platform—sign up for Confluent Cloud to stand up managed Kafka and Flink in minutes. 

Apache®, Apache Kafka®, Apache Flink®, Flink®, and the Flink logo are trademarks of the Apache Software Foundation in the United States and/or other countries. No endorsement by the Apache Software Foundation is implied by using these marks. All other trademarks are the property of their respective owners.