I built a real-time trading platform with an AI; then the clock started lying (faridsaid.com)

A developer harnessed AI to create a real-time price-matching platform for OTC interest-rate swaps, leveraging technologies like FastAPI, WebSocket, and SQLite. Although the application operated successfully for a month, issues arose when the session durations began to extend unpredictably, causing confusion among users. The problem stemmed from the timer mechanism, which wasn't inherently faulty but was affected by performance contention within the asyncio event loop, where multiple client requests could block each other. This incident underscores a critical lesson in asynchronous programming: being asynchronous doesn't guarantee concurrency. The developer resolved the issue by implementing a wall clock for timing and modifying the broadcasting method to allow concurrent sends to clients, improving overall system efficiency and reliability. By diagnosing the starvation of the timer due to resource contention and adjusting the architecture, the developer emphasizes how AI can assist but doesn't replace human reasoning and problem-solving, ultimately allowing engineers to focus on understanding complex issues. This case serves as a reminder of the intricacies of designing high-performance systems in the AI/ML landscape.