AI-Powered Fan Engagement and Content Personalization for Global Football Audiences

dfl_/_bundesliga · Media & Entertainment · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Sports media fan engagement and content operations

DFL/Bundesliga needed to personalize and localize football content for a global audience while making historical video archives searchable and reducing manual content-production work.

Constraints:

Fans span many countries, languages, clubs, and content preferences
Live match content must be generated with very low latency
Historical video archives lack enough metadata for automated discovery

Architecture

Bundesliga's AWS stack combines a near-real-time match data hub, Amazon Personalize for app feeds, intelligent metadata generation over S3 video archives with Transcribe/Bedrock/OpenSearch, localization workflows with Bedrock and DeepDub, Lambda-triggered live ticker generation, and AI Stories that use Bedrock plus Rekognition to transform long-form articles into mobile story content. Editorial workflows remain in the loop for live ticker and content quality.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  bundesliga-data-hub["bundesliga-data-hub"]
  amazon-personalize-feed["amazon-personalize-feed"]
  s3-video-archive["s3-video-archive"]
  transcribe-bedrock-metadata["transcribe-bedrock-metadata"]
  opensearch-media-index["opensearch-media-index"]
  localization-engine["localization-engine"]
  live-ticker-lambda["live-ticker-lambda"]
  bedrock-commentary-generator["bedrock-commentary-generator"]
  ticker-editor-workflow["ticker-editor-workflow"]
  stories-generation-pipeline["stories-generation-pipeline"]
  bundesliga-data-hub --> amazon-personalize-feed
  bundesliga-data-hub --> live-ticker-lambda
  s3-video-archive --> transcribe-bedrock-metadata
  s3-video-archive --> localization-engine
  transcribe-bedrock-metadata --> opensearch-media-index
  localization-engine --> s3-video-archive
  live-ticker-lambda --> bedrock-commentary-generator
  live-ticker-lambda --> ticker-editor-workflow
  stories-generation-pipeline --> bedrock-commentary-generator

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`bundesliga-data-hub`	Near-real-time match data and event foundation for downstream fan experiences.	`amazon-personalize-feed`, `live-ticker-lambda`
`amazon-personalize-feed`	Personalized app content recommendation layer for before, during, and after matches.	—
`s3-video-archive`	Video-file archive for historical and current media assets.	`transcribe-bedrock-metadata`, `localization-engine`
`transcribe-bedrock-metadata`	Metadata generation workflow that interprets commentary and video context.	`opensearch-media-index`
`opensearch-media-index`	Search index used for intelligent metadata queries over media files.	—
`localization-engine`	Speech-to-text/speech-to-speech localization workflow using Bedrock scripts and DeepDub voiceover.	`s3-video-archive`
`live-ticker-lambda`	Event-triggered Lambda that prompts Bedrock to draft live commentary.	`bedrock-commentary-generator`, `ticker-editor-workflow`
`bedrock-commentary-generator`	Generative model layer for live ticker commentary, scripts, metadata, and story generation.	—
`ticker-editor-workflow`	Human editor workflow that watches matches and publishes ticker entries in real time.	—
`stories-generation-pipeline`	Transforms existing articles into mobile story format with Bedrock prompts, Rekognition metadata, images, and upscaling.	`bedrock-commentary-generator`

Data flow

Match cameras and event feeds produce data into the Bundesliga data hub. Personalize uses fan behavior and content data to recommend app content. Archived videos in S3 are transcribed and enriched with Bedrock/Rekognition metadata, then indexed through OpenSearch for search and automation. Match events trigger Lambda prompts to Bedrock for live ticker drafts reviewed by editors, while article content is normalized, prompted, enriched with player metadata, paired with images, and published as AI Stories.

3 supporting quotes

“AWS lambda function is being triggered and that creates a prompt that asks Amazon Bedrock um via an API call to create a commentary on that event”transcript, stated

“content files, video files sitting in Amazon S3 and through Amazon transcribe and bedrock we would make sense of what a commentator is saying”transcript, stated

“Article normalization, prompt engineering, creation of metadata with AWS bed with Amazon Bedrock and Amazon recognition. Then it's image selection, upscaling and the story”transcript, stated

Why this architecture

Bundesliga needs multiple AI pipelines because fan engagement spans live events, archived media, localization, and personalized mobile consumption. Data infrastructure supplies factual match context, metadata makes huge video archives searchable, Lambda/Bedrock keeps live ticker latency low while editors retain control, and story generation repackages long-form articles for mobile attention patterns.

3 supporting quotes

“data is making a difference.”transcript, implied

“what is the right level of metadata that we need? What is what we say intelligent meta data.”transcript, stated

“we created the Bundesliga stories for the mobile app.”transcript, stated

Models used: Amazon Nova models · Claude 3.5 · DeepDub AI · Amazon Personalize

Scale

Bundesliga reaches 1 billion fans globally.
“the 1 billion fans Bundesliga has globally.”transcript, stated
The AWS partnership started in 2020.
“we have started the partnership in 2020 already.”transcript, stated
Match tracking collects player and ball positions 25 times per second and generates 3.6 million data points per game.
“track the position of the player and the ball 25 times per second generating data of 3.6 million data points per game.”transcript, stated
Skeletal tracking is expected to increase processing to 150 million data points per game.
“thanks to skeletal data, we actually process 150 million data points per game.”transcript, stated
The system also processes 1,600 event data points per game.
“we also process and create 1,600 event data.”transcript, stated
Personalization increased app usage by 20% and articles read by over 67%.
“increase the overall app usage by 20% and articles read by over 67%.”transcript, stated
Bundesliga has 75+ media partners and live/on-demand coverage in over 200 countries.
“We have 75 plus media partners. We have over 200 countries covering the Bundesliga live or on demand.”transcript, stated
Localization showed 75% processing-time reduction and 5x content output.
“numbers are quite staggering. 75% in processing time and five times the content created.”transcript, stated
Bundesliga produces 400 AI-generated stories per month, with 2x dwell time and 67% story retention.
3 supporting quotes
“So today Bundesliga produces 400 AI generated stories per month”transcript, stated
“it helped us to increase two times more the time spent in the app”transcript, stated
“stories was we had 67% story retention.”transcript, stated

Alternatives considered (roads not taken)

Manual-only live ticker commentary for every language and style

Rejected because: Lambda and Bedrock draft real-time ticker entries, while editors still watch and publish in real time.

“provide that to Bundesiga ticker editors who are watching the game in a real time and creating those tickers”transcript, stated

Searching video archives without dense metadata

Rejected because: The IGM workflow enriches video files with intelligent metadata to support search and automation.

“attach it to the video files and create a whole new level of metadata that we can use for automation.”transcript, stated

Leaving articles only in long-form format

Rejected because: Bundesliga Stories convert existing articles into a faster mobile story format.

“creating a new format of the content what we already have”transcript, stated

One-language/global content without localization

Rejected because: Research shows fans prefer content in their mother tongue, so Bundesliga localizes video and ticker output.

“We know from research that they prefer content in their mother”transcript, stated

Patterns

`latency-optimization` (real-time-match-commentary)

The live ticker generates goal context and goal cards within seven seconds, fast enough to beat broadcast delay.

“there's an automatically created a goal card with a player picture included and the context of the goal created and this is all done within the seven seconds which is super fast.”transcript, stated

`human-in-the-loop-gate` (editorial-review)

Generated live ticker commentary is delivered to Bundesliga ticker editors who are watching the match in real time.

“provide that to Bundesiga ticker editors who are watching the game in a real time and creating those tickers um in a real”transcript, stated

`batch-offline-processing` (archive-metadata-enrichment)

IGM processes archived S3 video files through transcription, Bedrock language understanding, visual recognition, and video-sequence context extraction.

“we would have um content files, video files sitting in Amazon S3 and through Amazon transcribe and bedrock we would make sense of what a commentator is saying in that video”transcript, stated

`tool-use-orchestration` (multimodal-content-pipeline)

The stories pipeline normalizes articles, prompts Bedrock, creates metadata with Bedrock and Rekognition, selects and upscales images, then pushes stories to the app.

“Article normalization, prompt engineering, creation of metadata with AWS bed with Amazon Bedrock and Amazon recognition. Then it's image selection, upscaling and the story is created and pushed to the Bundesliga app.”transcript, stated

`data-quality-loop` (localization-corrections)

The localization system learns from media-partner corrections to avoid repeating translation or voice-over errors.

“When customers or media partners identify errors in localized content, they can make corrections, and the system learns from these corrections to avoid repeating the same mistakes.”summary, stated

Tradeoffs

How to scale live match updates globally.

Chose AI-generated drafts with real-time editor workflow over manual-only commentary for every language and style.

Rationale: Automation provides speed, personalization, and scalability while editorial review preserves quality.

“ticker. It's a fully automated solution that is based on our match data and it's providing playbyplay textbased commentary that fills exactly that cap that gap. The gap is personalization and scalability.”transcript, stated

How to use existing long-form articles for mobile engagement.

Chose convert articles into AI-generated stories over leaving SEO articles only in their original long-form format.

Rationale: Stories repurpose existing content into the mobile format users spend more time consuming.

“switching the for the content we already have and switching to a new content which is a bonus stories in the app it helped us to increase two times more the time spent in the app”transcript, stated

Gotchas (what bit them)

Historical video archives are difficult to search without dense metadata.

Root cause: Older footage often lacks the detailed event, player, weather, and visual context metadata available for modern matches.

Resolution: Bundesliga layers AI-generated transcription, visual segmentation, and contextual video metadata onto official match data.

“these video sequences we can then use with 12 labs to understand the context. Is it raining in that video? Is there someone holding up their scarf with uh a club name for example?”transcript, stated

Voice localization still has quality gaps compared with original content.

Root cause: Speech-to-speech generation remains imperfect despite large improvements.

Resolution: The system delivers speed and volume gains while allowing customers and partners to adjust generated content.

“The presenters acknowledge that quality differences remain between the original and generated content, showing appropriate transparency about current limitations.”summary, stated

Audience-specific commentary style is still an experiment.

Root cause: Different demographics may respond differently to professional, casual, or youth-oriented tone.

Resolution: Bundesliga uses prompt-engineered style variants and learns from engagement response.

“They admit uncertainty about whether the "bro code" style commentary will resonate with younger audiences, framing it as an experiment worth pursuing.”summary, stated

Evaluation

Bundesliga evaluates the systems through app usage, article reads, localization throughput, live-generation latency, story retention, app dwell time, and efficiency gains.

Metrics: overall app usage · articles read · processing time · content volume · story retention · app dwell time · generation latency

“through Amazon Personalize, we're actually able to increase the overall app usage by 20% and articles read by over 67%.”transcript, stated

“The result of Bundesiga stories was we had 67% story retention. So the users, the fans who watch the stories, they stay much longer in the app”transcript, stated

Outcomes

Personalization increased app usage by 20% and articles read by 67%.
Localization reduced processing time by 75% and created five times more content.
Bundesliga produces about 400 AI-generated stories per month, with 67% story retention and doubled app dwell time.

The summary describes production systems for a one-billion-fan global audience, 200+ countries, 75+ media partners, and 9+ petabytes of archival footage.

“75% in processing time and five times the content created.”transcript, stated

“today Bundesliga produces 400 AI generated stories per month”transcript, stated