FloVision Builds AI-powered Food Processing Operations with FiftyOne.

Introduction: FloVision

FloVision builds AI-powered inspection and measurement systems for food processing operations. Their solution helps meat production facilities improve yield accuracy, ensure product quality, and improve labor performance through real-time object detection, tracking, and measurement.

Food packaging conveyor belts and workstations are connected with compact sensors that scan every product as it moves through the production line. FloVision has processed over 23 million kilograms of food and helped processors globally improve yield by 1.5%. Their mission is to reduce greenhouse gas emissions by 1% globally by helping processors reduce waste and make better use of the resources that feed communities every day.

"Our mission is to create a financially and environmentally sustainable future for the food industry. We're equipping food processors with AI-assisted automation and analytics that streamline production and increase yield and quality across the line." — Rian McDonnell, Founder & CEO, FloVision

Delivering these results requires FloVision's ML team to continuously develop, train, and iterate on computer vision models across massive volumes of production images and video.

FloVision's analytics dashboard gives food processors real-time visibility into yield, quality, and operator performance.

The Challenge: Too much data, but too little visibility

FloVision's ML team develops and maintains vision models for a variety of applications, including:

A yield estimation system that detects regions of interest on beef/chicken carcasses and measures meat thickness to assess yield against customer targets.
A box label verification system that counts, classifies, and validates packaged meat being prepared for shipment

Both use cases depend on continuous, high-quality video data from multi-sensor production environments, and require tight iteration cycles to keep models performing to spec.

FloVision collects hundreds of thousands of videos during production, but not all are useful for model training. The FloVision team faced two big challenges:

There wasn’t a fast way to identify which video footage was worth using for model development.
Finding edge cases took even longer. When model performance dipped, diagnosing the root cause meant manually sifting through data for days or sometimes an entire week before the team could pinpoint what was wrong.

With ML work spread across tools and one-off scripts, there was no consistent foundation for the team to build on together.

"Yes, you have the data, tons of it, but there was no easy way for us to filter through large swaths of it. We could only discern it from tags or classes, nothing more visual, nothing deeper. It was a big time sink." — Terrance Whitehurst, ML Researcher, FloVision

Why FiftyOne

As data volumes grew and use cases expanded to video-heavy production-line workflows, FloVision needed deeper visibility into their data and model performance.

Visual exploration at scale: Video thumbnails made it faster to identify usable footage at a glance.

Embedding-based edge case detection: By visualizing image clusters, the team could surface underperforming model behavior such as edge cases or outlier behavior, in hours instead of days.

Extensible framework. Engineers are able to build reusable operations and custom functionality inside FiftyOne's UI that the entire team can access and invoke.

Enterprise deployment flexibility. A single managed instance replaced fragmented infrastructure and enabled real collaboration across shared datasets.

How FloVision Uses FiftyOne

FloVision uses FiftyOne as the operational hub of their ML pipeline, bringing together data ingestion from connected AWS S3 buckets, visualization and curation, auto-labeling, and expert-level collaborative annotation, and model training and evaluation into one repeatable workflow.

For their packing line verification product, annotators review auto-labeled frames in shared FiftyOne datasets. With side-by-side top and side camera views, labelers can cross-reference and correct annotations without switching tools. Good frames flow automatically into training datasets; bad ones are flagged for relabeling. Curated data is then fed into the AI training infrastructure on a set schedule or on demand.

For their poultry yield estimation product—which generates millions of images per day—the team uses FiftyOne's embedding visualization to identify the most unique image clusters. This has become their primary method for diagnosing model performance dips and surfacing edge cases before they affect production.

"The team has built a growing library of custom FiftyOne extensions that any team member can invoke from the UI — from data transformation routines to custom model runners. Rather than things being on disparate GitHub repos that we may or may not have fully fleshed out, the team's operational tooling now lives in one place, stable and accessible to everyone." Terrance Whitehurst, ML Researcher, FloVision

FloVision Nano delivers real-time operator guidance and yield analysis directly on the production floor.

Results: Faster data workflows helped FloVision deliver a market-differentiating packing line verification system

FloVision's automated packing line verification product required curating over 200,000 videos. FiftyOne's visual pipeline made it possible to move through that volume at the pace the project demanded.

"We would not have been able to get through this massive product milestone for our project in the timeline we needed to without FiftyOne." — Terrance Whitehurst, ML Researcher, FloVision

7x faster model edge case detection: Using FiftyOne's embedding visualization to identify unique image clusters, the team now surfaces model edge cases and performance anomalies the same day they investigate; down from nearly a week of manual data review.

"Some of our dips in model performance were because of edge cases. With FiftyOne, we were able to catch them the same day we analyzed model performance, something that would have maybe taken a week otherwise. " — Terrance Whitehurst, ML Researcher, FloVision

Higher team-wide productivity through shared operations: Automated plugins for data ingestion, temporal visualization, labeling, and training pipeline orchestration have improved development productivity. The team can get more data into models faster. This has translated to continued improvements in model performance across both product lines.

"The fact that we're able to develop pretty much anything we want with our data and put it into reusable components in the UI has been incredibly helpful for repeatability, stability, and team collaboration. Every time someone creates something, it raises the floor of what every single person can do every time they open FiftyOne." — Terrance Whitehurst, ML Researcher, FloVision

7x Faster Model Analysis: How FloVision Is Reducing Millions of Kilograms of Food Waste Using FiftyOne