Dobble Agent.png

Idea

Dobble Agent - our award winning Dobble implementation with a side serving of open source agentic implementation.

Demonstrate:

Can we combine Scott’s low LLM and Chris’ Dobble VL projects as well into this?

What does it actually do?

This is the tricky bit. It’s playing dobble but then there needs to be some kind of agentic ‘action’ either before, during or after the game. Some thoughts:

Business Case

Dobble was one of our most successful R&D projects to date. The whole team came together (5 engineers at the time) and won the ZenML month of MLOps competition with our entry.

It showcased great MLOps (at the time, November 2022) with an interesting use case of computer vision on edge.

ChatGPT came out about a week after we won the competition and since then we’ve done a lot of work on LLMs. We got in there fairly early in demonstrating how to build an LLM implementation using open source (with MindGPT) and that led to a number of customer projects off the back of that work.

Over the past 6 months or so there has been a lot of talk about agentic AI being ‘the next big thing’ but uncertainty over what it actually is AND how to engineer it properly have been the source of much conversation.

Recently however, some open standards have started to emerge. Notably the Model Context Protocol (MCP) from Anthropic came in November 2024 but in the last few weeks it has really started to pickup in terms of demonstrable implementations.

ZenML have (as usual) done a really nice implementation which shows how to interact with their MLOps pipelines using natural language.