A talk by Dario Cannone from Faculty.
In this talk, I will share my experience of building an LLM-based agent for a UK fintech company and deploying it in a complex production ecosystem.
While agents powered by Large Language Models (LLMs) are often hyped for their potential to 'think' and solve complex tasks, they can often result in unpredictable interaction loops that lead to unexpected outcomes and loss of control over the system. This is particularly worrying when, as in our case, the LLM is given the power to interrogate internal company data.
There is a way to mitigate these risks, which is to build a much simpler LLM agent and force it to solve a task following a pre-determined chain of simpler subtasks, sacrificing generalisability for control. In this talk, we will discuss the development of such an agent, as well as the advantages and disadvantages of our choices.
Technical Level: Technical practitioner
This session was part of the Data Science Festival MayDay event 2024. Find out more at datasciencefestival.com/event...
The Data Science Festival is the place for data-driven people to come together, share cutting-edge ideas, and solve real-world problems. We run monthly events, meet-ups, and the biggest free-to-attend data festivals in the UK. Join the community at datasciencefestival.com/