Alegion teaches AI to keep its eye on the ball

A soccer ball is an inconvenient object. It is small, it moves quickly, and it has a habit of disappearing behind people whose profession involves getting in its way. For a computer watching a match, the difficult question is often quite ordinary: where did it go?

The model had a labeling problem

Sportlogiq, the Montreal sports analytics company, had a soccer model with persistent errors. Alegion’s published case study traces the problem to unreliable annotations. Michael Jamieson, an AI research manager, also faced a management question: should his team run a separate labeling workforce? An internal cost-benefit analysis persuaded it to outsource.

“We did a cost-benefit analysis and decided that doing it internally didn’t make sense.”

Michael Jamieson · Sportlogiq

Alegion split ball and player labeling into separate workflows, tested the interfaces, and used manual labels on every tenth frame with tracking between them. Its case study reports 4,116,563 annotations delivered in 77 days and a 70% reduction in model error. Those are customer-project results, rather than a promise for every dataset.

The useful lesson is to test a vendor on difficult scenes before handing over the entire job. A proof of concept gives both sides something concrete to argue about: the actual labels.

A crowd becomes a production line

Alegion’s beginnings were broader than sports. In November 2012, it publicly launched at the first AWS re:Invent conference. Amazon’s announcement described an enterprise workflow designer built around Mechanical Turk and Simple Workflow Service. A business could put crowd labor into a larger process instead of managing a collection of disconnected tasks.

Co-founders Nathaniel Gates and Joel Simpson were working on a coordination problem. Finding people was one part of it. Specifying their work, routing it, and deciding whether the result deserved acceptance were the other parts. As machine learning demanded labeled examples, that coordination became a business in its own right.

The financing followed. Alegion announced a $3.6 million Series A led by RHS Investments in June 2017; a further $12 million Series A-2 was reported in August 2019. The 2017 announcement described training datasets, human-scored validation, and exception processing. Humans had jobs before a model entered production and after it encountered something awkward.

In November 2020, Alegion launched Control, a self-service video annotation product supporting 4K and long-form footage. Then-CEO David Mather said clients wanted direct control as their labeling needs changed. The launch put the company’s software into customers’ hands, adding a SaaS offering to its managed work.

Seventeen joints, one factory floor

Invisible AI encountered a different kind of missing information. Its manufacturing vision system began with the public COCO pose dataset, but general-purpose data did not adequately capture workers partly obscured by car components, or the lighting and angles of an automotive factory. The setting mattered as much as the human figure.

Alegion’s case study describes a deliberately specific rulebook: annotate 17 body keypoints, include people at least 50% visible and larger than 200 pixels, and manually label every third frame. Software tracking supplied the intervening frames. The project produced 3,432,000 labels in 11 months; the reported weighted limb F1 score moved from 0.74 to 0.79.

Every third frame here, every tenth in soccer. That difference is instructive. A sampling interval is a decision about a particular task. Copying the number without checking the footage would miss the point.

What the customer actually buys

Alegion now describes an end-to-end service: define the annotation strategy, configure workflows, recruit and train workers, label the data, calibrate the project, and review quality. It covers images, video, audio, and text. Customers can seek bounding boxes, detailed outlines, pixel-level segmentation, transcriptions, or labels for entities and sentiment in language.

The commercial proposition is a managed production process. An AI team supplies requirements and data; Alegion organizes delivery around quality, budget, and time. For a buyer, the sensible cost comparison includes engineering time, supervision, retraining, and rejected work alongside the price of a label. A cheap label becomes less charming when somebody has to repair it.

Alternatives include managed providers such as CloudFactory and iMerit, annotation platforms such as Labelbox, and an internal operation. Alegion’s pitch rests on combining workflow software with people who manage the work. Conditional logic, multi-stage tasks, and quality routing help explain its interest in complicated assignments.

Its insurance examples include outlining fallen trees on damaged roofs and classifying vehicle damage. In security, it describes distinguishing animals from people and relating hands, scanners, and groceries at checkout. The common thread is a judgment that needs a precise definition before a machine can learn it.

Alegion’s process starts with representative samples and a proof of concept, then uses worker qualification, scoring, review, and feedback. Buyers can copy that sequence. It loses its value if the sample misses the operating environment, the rules remain disputed, or performance is never checked on fresh examples. Increasing annotation volume cannot settle an unresolved definition.

The people remain in the picture

The workforce is also part of the product. Alegion says it has impact centers in Uganda and Egypt. In 2021, it announced a Malaysian partnership with Yayasan Peneraju targeting training for 1,600 labelers. Its stated ambition is dignified work and useful skills; those commitments deserve attention alongside delivery metrics.

In 2024, Gates announced the SanctifAI transition, tying it to agent workflows with human participation. Today, SanctifAI lists Alegion Data Services in its family and Stephanie Lee as its managing director. The data business continues within a broader set of AI activities.

That makes the soccer ball a useful place to finish. Someone must specify what counts as finding it, decide which examples matter, and check whether the system learned the intended lesson. Alegion sells the organization of that work. The machine gets the highlight reel; the labels carry the plot.