One Tool to Rule them all: Building a Fast AI Agent for Real-Time Response
When your goal is to make an agent fast enough to explore millions of data points and find the true cause of incidents in minutes, the way you engineer agents fundamentally changes.
Come with me as we explore how we built a fast and reliable agent for live risk limit breaches with only one tool: a Python interpreter. We’ll cover the kind of custom backpressure and speedups you gain while in an interpreter, some whacky checks you can do with ASTs and the design trade-offs you face when building agents in this manner
Ryan Samarakoon
Ryan Samarakoon is the primary software engineer/contributor for the APAC Risk Team’s AI strategy. He has worked at Optiver for 3 years and has been building production grade AI solutions for Risk ever since agentic AI took off.
His work in the AI space is unique as it works close to the high risk, low forgiveness intersection of trading and risk management, where speed and accuracy are key.