A presentation at Devoxx UK in in London, UK by Mark Needham
Game, Set, Match Transforming Live Sports with AI-Driven Commentary Dunith Danushka (Redpanda Data) Mark Needham (ClickHouse)
Dunith Danushka DevRel @ Redpanda Mark Needham Product @ ClickHouse
Can we create an AI Copilot to help live text writers?
What are we going to build?
The flow of events Window queries on ClickHouse and pass the results to OpenAI
What is Redpanda?
Redpanda is a Kafka API compatible streaming data platform ● ● ● ● Not a Kafka fork! Kafka rewritten in C++ Identical read/write interfaces as Kafka Designed for modern hardware
Simple to deploy, use and manage Single binary Kafka-compatible APIs Easy Day 2 Ops Dev-friendly interface © 2023 REDPANDA DATA
rpk Redpanda’s command line interface (CLI) utility. Check health of cluster rpk cluster health Create a topic rpk topic create my-topic -p 5 List topics rpk topic list Describe a topic rpk topic describe
Redpanda Demo
What is ClickHouse?
What is ClickHouse? Open Source Distributed Column-Oriented OLAP Database Developed since 2009 Files per column Replication Analytics use cases OSS 2016 Vectorized query execution Sharding Aggregations 34k+ Github stars Optimised for aggregations Multi-master Visualizations 1k+ contributors Sorting and indexing Cross-region Mostly immutable data 500+ releases Background merges
Row Oriented vs Column Oriented Row Oriented location ts temperature wind_speed humidity Aberystwyth 2022-01-01 00:00:00 14 21 79 Blackpool 2022-01-01 00:20:00 13 9 82 Column Oriented location ts temperature wind_speed humidity Aberystwyth 2022-01-01 00:00:00 14 21 79 Blackpool 2022-01-01 00:20:00 13 9 82
Vectorised Query Execution Process rows sequentially Process chunks of values
Flavours of ClickHouse chdb ClickHouse Local ClickHouse Server
ClickHouse Demo
What is Streamlit?
What is Streamlit? Streamlit turns data scripts into shareable web apps in minutes. All in pure Python. No front-end experience required.
Streamlit Hello World
Building the AI Copilot
The flow of events Window queries on ClickHouse and pass the results to OpenAI
Generated Commentary Demo
How do Large Language Models work? Prompt LLM
Ingesting context into the prompt Instructions Prompt Context Sports events feed
Sports events feed
LLM Code
Pulling events from ClickHouse
Generated text events
Retrieval queries: Recent points
Retrieval queries: Last completed game
Serving the live commentary Multiple consumers on the Redpanda topic FastAPI server that renders SSE events Post messages to Twitter API points livetext
Live Commentary Demo
Future Ideas
How can we extend this work? ● ● ● ● ● ● Automatic summaries every <x> seconds A Copilot that has access to fine-grained statistics Text to SQL so that the writer can ask questions of the data Can we use more batch data? Store the generated commentary for later analysis/tweaking of the prompt Compare the generated commentary with what the analyst chooses to publish
Is it only for sports? Weʼll be focusing on sports, but you could use it for any of the following problems: ● ● ● ● ● Live auctions Weather updates Local traffic reporting Current location of food delivery … Any use case where you have events that you want to summarise into a more readable format.
Thanks and Questions github.com/mneedham/devoxx-ai-sports-commentary www.linkedin.com/in/dunithd dunith.medium.com www.linkedin.com/in/markhneedham youtube.com/@LearnDataWithMark
We are both big fans of the live text commentary the BBC provides for sports like football, tennis, rugby, cricket, and more. While there are a lot of novel observations in the commentary, there is also a lot that effectively summarises what just happened.
Wouldn’t it be cool if the commentator could have a Co-Pilot who can make the process more efficient?
This session will introduce an AI co-pilot for sports commentary based on Redpanda, ClickHouse, and a large language model. A stream of events will be fed into RedPanda, and we’ll capture a window of those events in game-by-game and/or time-period buckets using ClickHouse.
We’ll then send the LLM the events that have just happened along with queries on historical data, from which it can generate suggested text commentary. The commentator can then decide whether to use the Co-Pilot’s suggestion, edit the suggestion, or just go along with their own version.
The following resources were mentioned during the presentation or are useful additional information.
for free. You
can too.