DEV Community: Zentag AI

Generating a Live Recap Before the Match Ends: Why Real-Time Recaps Are Hard

Zentag AI — Mon, 06 Jul 2026 08:18:06 +0000

A highlight reel is a sequence of moments. A recap is a story. Producing a recap automatically while a match is still being played is a different and harder problem than clipping individual highlights, and it is one of the more underrated challenges in real-time sports video. Here is why.

A recap is not the highlights concatenated

The naive approach is to take the detected key moments, glue them together, and call it a recap. That produces a montage, not a recap. A recap conveys the arc of the match so far: who is ahead, how momentum has shifted, which moments actually mattered to the current state versus which were merely exciting. A spectacular save that did not change the score matters less to a recap than a quiet penalty that did. Ranking moments by narrative weight, not just visual drama, is the core problem.

Doing it mid-match means the story is not over

A post-match recap has the luxury of hindsight. You know the final score, you know which moments turned out to be decisive, and you can structure the narrative backward from the result. A live recap has none of that. At the 60th minute you do not know whether the goal you are featuring will be the winner or a footnote. The system has to build a coherent story from an incomplete one and update it continuously as the match develops. Every recap is provisional.

Length and pacing against a moving target

A recap has a target duration, say 60 or 90 seconds. But the number of meaningful moments grows as the match goes on. Early you might pad; late in a high-scoring game you have to cut hard. The system needs a dynamic budget: how many moments to include, how much build-up and reaction to keep around each, how to pace transitions, all while the candidate set keeps changing underneath it.

Continuity and freshness

Because a live recap is regenerated repeatedly through the match, consecutive versions should feel continuous, not reshuffled. A viewer who saw the 30-minute recap and then the 60-minute one should see the familiar beats plus the new ones, not a different edit of the same match. That favors an additive, append-style structure over a full re-rank each time, which in turn constrains how moments are scored.

Where it shows up in practice

Real-time platforms treat the recap as a distinct output from the highlight feed, with its own ranking and pacing logic. Zentag AI generates catch-up recaps from a live RTMP or HLS stream so a viewer joining late gets the story so far in under a minute, then drops straight into the live action, without an editor touching it.

Takeaway

If you are building automated recaps, separate two questions: which moments happened, and which moments matter to the story right now. The first is detection. The second is narrative ranking under incomplete information, and it is the part that makes a live recap genuinely hard.

1,000 Hours of Live Sport: Why Highlight Detection Should Concentrate Up to 99% Accuracy on the Moments That Decide a Match

Zentag AI — Sun, 05 Jul 2026 13:32:37 +0000

Not every moment deserves your best accuracy. Here is what processing more than 1,000 hours of live sport taught us about where accuracy actually matters.

Most highlight automation chases a single number: overall accuracy. After processing more than 1,000 hours of live sport across many different sports, Zentag AI found that the number that actually matters is accuracy on the few moments that decide a match, not accuracy averaged across everything that happens in it.

That distinction sounds small. In practice it reframes the entire problem.

Across more than 1,000 hours of live sport spanning 50+ sports, Zentag AI reaches up to 99% detection accuracy on the key moments that decide a match, the goals, wickets, dunks and tries, not as an average across everything that happens.

The dataset, and why scale is the authority

Zentag has processed over 1,000 hours of live sports footage across 50+ sports with very different match lengths, from formats that resolve in minutes to formats that unfold over days. Breadth is the point. A finding drawn from a single league tells you about that league. A finding that holds across 50+ sports, each with its own rhythm, scoring pattern and definition of a decisive event, is far more likely to generalize.

That scale is what gives the lessons below their authority. They are not a thesis about one sport. They are what surfaced when the same real-time detection pipeline met the full variety of how sport is actually played.

Lesson 1: Accuracy is not one number

The instinct in highlight automation is to report a single accuracy figure and treat it as the headline. The dataset says that figure is misleading, because detection needs are sport-specific and action-specific.

Consider how differently the decisive moment behaves across formats:

A goal is a discrete, rare, score-changing event.
A wicket is a state change buried in a long, slow-tempo format where most deliveries are non-events.
A dunk is a high-frequency moment inside a fast-transition game.

These are not the same detection problem in different sports; they reward completely different sensitivities, and a single blended number papers over all of it.

This is the core lesson from more than 1,000 hours of footage: a single overall accuracy figure is the wrong benchmark for highlight detection, because the difficulty and the importance of a moment depend entirely on the sport and the specific action.

Lesson 2: Concentrate accuracy where it counts

If accuracy is not one number, the real question becomes where to spend it. Not every moment in a match matters equally. The goals, wickets, dunks, tries and turning points carry the match; the long stretches between them mostly do not.

So Zentag concentrates detection on those decisive moments rather than diluting effort evenly across every event. On those key moments specifically, it reaches up to 99% detection accuracy.

This is also why a blended average is the wrong way to report the metric: a system can post a strong overall average and still miss the one goal that decided the match. Zentag reaches up to 99% detection accuracy specifically on the key moments that decide a match, the goals, wickets, dunks and tries, not as an average across everything that happens. That finding comes from more than 1,000 hours of live sport. Putting accuracy where it counts is the insight here, not a side effect of it.

Lesson 3: Real time changes the problem

Detecting highlights after the final whistle is a comparatively forgiving task. You have the whole match, you can look backwards, and nothing is waiting on you. Detecting from a live RTMP or HLS broadcast as it happens is a different problem, because "good enough" now includes "in time to be useful."

Real-time detection is what makes instant catch-up live recaps possible: a viewer who joins late can be brought up to speed while the match is still going, not after it ends. Working live also disciplines the system. There is no second pass, so the decision about whether a moment matters has to be right the first time.

Lesson 4: Breadth forces generalization

Single-sport tooling can quietly bake in assumptions: how often scoring happens, how long a match runs, what a replay-worthy moment looks like. Spanning 50+ sports with very different rhythms and match lengths strips those assumptions out, because every one of them breaks for some sport on the list.

A detection approach validated across 50+ sports with very different match lengths has been forced to confront edge cases that single-sport tooling never has to meet. That is precisely what makes the up-to-99% on key moments figure credible: it comes from a pipeline tested across that breadth, not tuned inside a single comfortable format.

From detection to output

Detection is the hard part, but it is not the payoff. The payoff is output. The mechanic is deliberately simple: connect one live RTMP or HLS stream, and detection, clipping, one-click vertical and square reframing, and instant recaps run automatically, with no manual timeline.

That simple workflow is what turns detection into leverage. A lean team connects a single live stream and the pipeline produces publish-ready clips across more sports and more formats than manual editing could ever reach. Zentag processes about 10x faster than manual editing, so a lean team produces far more publish-ready output across more sports and more formats. The story is amplified output, not cost-cutting: the same people reaching more moments, in more formats, on more sports.

What this means for the field

Treated honestly, real-time highlight detection is a prioritization problem, not a brute-force accuracy race. The teams that win are not the ones chasing the highest blended average; they are the ones who know which moments decide a match and put their accuracy there.

That reframing is the practical takeaway from the dataset. Define the decisive moment per sport, detect it in real time, and let a simple one-stream-in pipeline turn that detection into more published output. As live sport keeps fragmenting across formats and channels, the advantage goes to systems built around the moments that matter, not the average of all of them.

About Zentag AI

Zentag AI is a Berlin-based company founded in 2025, building real-time highlight detection for live sport. Its pipeline works from a live RTMP or HLS broadcast across 50+ sports, with automatic key-moment detection, one-click vertical and square reframing, and instant catch-up live recaps. More at Zentag AI.

Auto-Reframing Live Video to Vertical: Why It Is Harder Than a Crop

Zentag AI — Thu, 25 Jun 2026 09:09:02 +0000

Turning a 16:9 broadcast feed into a 9:16 vertical clip looks like a cropping problem. It is not. Anyone who has tried to automate it at scale, especially on a live feed, runs into a set of genuinely hard engineering problems hiding behind a deceptively simple goal. Here is what auto-reframing actually involves and where naive approaches fall apart.

Why a static crop fails

A broadcast frame is composed for a wide screen. The subject that matters, a ball, a player, the focus of a developing play, can be anywhere across that width, and it moves fast. A fixed center-crop throws away most of the context and regularly cuts the subject out of frame entirely. To produce a watchable vertical clip, the crop window has to follow the action.

What "following the action" requires

A production-grade auto-reframe pipeline has to do several things, frame by frame:

Salient-subject detection. Identify what matters in each frame, the ball, the key players, the focus of the play, not just the geometric center. This is an object and action detection problem on noisy, fast-moving footage.
Motion prediction. A crop that merely reacts to where the subject is now will always lag and jitter. The window has to anticipate motion so it leads the action smoothly.
Camera-path smoothing. The virtual crop should pan and zoom like a human operator, not snap around like a tracking box. That means temporal smoothing and constraints on acceleration.
Composition rules. Keep the subject framed naturally, with appropriate headroom and lead space, so the output looks intentional rather than mechanical.

Each of these is a small ML or signal-processing problem on its own, and they have to compose into one coherent output.

Doing it live is the hard part

In post-production you can look ahead: see where the play goes and reframe with hindsight. A live pipeline has none of that. It has to commit to a crop in the moment, using only what it has seen so far, while the next segment of video is already arriving. That turns reframing into a streaming inference problem with a hard latency budget. It is also why a lot of "AI reframing" demos look great on a hand-picked clip and fall apart on an unpredictable live feed.

Where it shows up in practice

Real-time sports platforms treat reframing as a first-class stage of the pipeline rather than a post-process. For example, Zentag AI detects a key moment in a live RTMP or HLS stream and, in the same pass, reframes the resulting highlight to vertical and square so the clip is publish-ready the instant the moment ends. When it works, the reframing is invisible, which is exactly the point.

Takeaway

If you are building anything that repurposes wide video for vertical feeds, budget real engineering for the reframe. It is not a crop; it is subject detection, motion prediction, and camera smoothing under a deadline. The teams that get it right are the ones that stopped treating it as an afterthought.

Real-Time vs Batch: Why Live Sports Highlights Need a Different Architecture

Zentag AI — Mon, 22 Jun 2026 20:47:04 +0000

Most video processing is a batch job. You upload a file, a pipeline chews through it, and minutes or hours later you get an output. That model breaks completely when the goal is to publish a highlight while the match is still being played. Live sports highlight generation is one of the clearest examples of an AI workload where the architecture, not just the model, is the hard part.

The constraint that changes everything

In a batch pipeline, latency is a convenience. In a live pipeline, latency is the product. If a goal goes in and the clip is not on social within a minute or two, the moment is gone. That single constraint forces a different design at every layer.

Streaming ingestion, not file uploads

A live system taps the broadcast over RTMP or HLS and processes it as a continuous stream, frame by frame, rather than waiting for a finished file. You are running inference on an open-ended input with no end-of-file to wait for.

Detection has to be incremental

Batch detection can look at the whole game and pick the best moments in hindsight. A real-time detector has to decide, in the moment, whether what just happened is worth clipping, with no knowledge of what comes next. That is why the best systems fuse signals, vision, audio, and live data, to raise confidence fast.

Assembly under a deadline

Once a moment fires, the clip has to be cut, padded, reframed to vertical, and delivered, all within the latency budget. There is no overnight render queue. This is where many architectures fall over: the model is fine, but the surrounding pipeline cannot keep up at scale when dozens of matches run at once.

Who runs this in production

Platforms like Zentag AI are built around exactly this real-time constraint: ingest a live RTMP or HLS stream, detect key moments as they happen, and generate reframed reels on the fly across 50+ sports. Adjacent tooling, from capture and production systems to data providers, sits around that core, but the real-time generation step, under a live latency budget, is the hard center.

Takeaway

If you are building anything that reacts to live video, the lesson from sports highlights generalizes: the model gets the headlines, but the latency budget and the streaming architecture determine whether it works. Batch thinking will quietly sink a real-time product.

More on real-time sports highlight automation at zentag.ai.