Sports data analytics for technique

Genius Sports just posted Q2 revenue of $196 million, up 65% year on year, driven by betting and media contracts that rely on structured event data (MarketBeat). That number tells you where the money is in sports data analytics today, and it also tells you where the ceiling is. Betting feeds are a solved category. Technique analysis is not.

The next wave of sports data analytics is about what happens between the events. Not the score, not the shot count, but the biomechanics of the swing itself. This is the space our client SportAI works in, and it is a much harder engineering problem than tagging goals.

Why technique is the hard problem in data analytics and sports

Event data is cheap. A dozen vendors will sell you a feed of every shot, pass, and point in a professional match. The signal is coarse but reliable, and the pipelines have been battle tested for a decade.

Technique data is different. It requires extracting joint positions, racket angles, ball trajectories, and swing paths from ordinary video, then interpreting them against a model of what a good stroke looks like. The engineering stack sits at the intersection of computer vision, pose estimation, physics informed models, and domain expertise from coaches.

This is where the interesting work in data analytics and sports is happening right now. It is also where most product teams get stuck, because the talent pool is narrow and the iteration loop is slow.

What the pipeline actually looks like

A production grade technique analysis pipeline typically involves:

• Video ingestion from consumer devices, often at inconsistent frame rates and angles.
• Player and ball detection under partial occlusion, glare, and cluttered backgrounds.
• Pose estimation, usually a fine tuned variant of a transformer or CNN based model.
• Temporal smoothing to remove jitter across frames.
• Biomechanical modelling to convert 2D or 3D keypoints into meaningful metrics like racket head speed or contact point.
• A feedback layer that turns raw metrics into coaching cues an athlete can actually use.

Each stage has its own failure modes, and none of them are solved off the shelf.

How Devspace approaches sports data analytics

We work with SportAI, an Oslo based B2B sports tech company applying computer vision and machine learning to real time technique analysis across racket sports and beyond. They needed flexible access to rare engineering talent, specifically senior AI and computer vision engineers who could ship into a live product rather than write papers.

We placed those engineers directly into the SportAI team through our remote development team model. They work inside SportAI's stack, sprint, and code review process. This is deliberate. In a domain this specialised, context transfer eats months if you run engagements as arms length statements of work.

The staffing profile that actually works

For sports data analytics products, we look for a specific combination that is rare in the general developer market:

1. Computer vision engineers with production experience, not just research background.
2. ML engineers who understand latency budgets, because coaching feedback that arrives ten seconds late is useless.
3. Backend engineers comfortable with video pipelines and GPU scheduling.
4. At least one person who has shipped a mobile SDK, because that is usually where the video comes from.

Generalists will not get you there. Neither will pure researchers. The bar is people who can move from a pose estimation paper to a shipped feature inside a sprint.

What real time technique analysis unlocks

Once the pipeline is stable, the product surface expands quickly. A racket sports platform can move from post match summaries to live drills, from single player analysis to opponent scouting, from consumer coaching to federation grade analytics.

This is where the commercial case gets interesting. Betting data monetises through B2B feeds. Technique data monetises through subscriptions, federation contracts, equipment brand partnerships, and licensing to broadcasters who want richer overlays. The unit economics look more like SaaS than like sports data licensing.

The teams building here are also feeder platforms for adjacent categories. The same pipeline that analyses a tennis serve can be adapted to golf, baseball, cricket, and eventually team sports. Racket sports is a good starting wedge because the environment is controlled, the players are stationary at contact, and the biomechanics are well studied.

Sports data analytics internships and building a talent pipeline

A quick note on talent, because we get asked about this. Search volume around sports data analytics internships is real, and the field is producing more graduates every year with the right foundation in statistics and ML.

Internships are a reasonable way to build a research bench, but they do not replace senior engineers on a live product. The gap between someone who can train a pose estimation model on a clean dataset and someone who can debug why the model drops the racket at frame 47 of a real user video is enormous. Product teams that lean too heavily on early career talent tend to ship demos, not products.

The blended model works better: senior engineers own the pipeline architecture and the hard debugging, interns and juniors work on data labelling, evaluation harnesses, and specific features under supervision.

A framework for evaluating a sports data analytics build

If you are a CTO or founder scoping a technique analysis product, here is the checklist we use when we start an engagement:

1. Data reality check. How much labelled video do you actually have, and how representative is it of end user conditions? If the answer is "we will collect it as we go", the project is 12 months longer than you think.
2. Latency budget. Define the end to end latency target before you pick models. Real time coaching is a different product from post session review.
3. Metric validity. Do you have a coach or biomechanist who can tell you whether your metrics are actually meaningful? Without this, you are shipping numbers, not insight.
4. Edge vs cloud split. Where does inference run? This decision cascades into hardware costs, battery life, and privacy posture.
5. Evaluation harness. Before you optimise, can you measure? A rigorous eval set that mirrors production video is the single highest leverage investment.
6. Staffing bar. Are your senior engineers senior enough to say no to the wrong architecture in month one?

Get these six right and you have a product. Get any two wrong and you have a research project.

Where this goes next

The money in sports data analytics is still concentrated in betting and media rights, and the Genius Sports numbers make that clear. But the interesting engineering, and the interesting long term margins, are moving toward technique, biomechanics, and personalised coaching.

That shift favours teams who can combine AI and data capability with the discipline to ship into consumer devices. It punishes teams who treat computer vision as a research exercise rather than a product one.

If you are building in this space and need senior computer vision or ML engineers embedded in your team within a few weeks, that is the work we do. Racket sports today, everything else soon after.

Photo by John Fornander on Unsplash