ClosePitch is an AI-powered sales productivity platform designed to help sales professionals write custom pitches. Building and launching it as a solo engineer was an intense exercise in balancing speed and maintainability.
One of the major lessons was that LLMs are highly unpredictable. Early versions frequently returned malformed JSON, which broke the Flutter front-end parsing logic. The solution was implementing validation checks in the Go API backend, dynamically adjusting parameters and retrying on failure before returning payloads.
Another lesson was caching. Querying DB servers for every pitch preview is expensive. Implementing Redis to cache active session states and prompt templates reduced server response times by 60%.
Finally, code structure matters. De-coupling presentation widgets from repository connections allowed rapid UI refactoring when updating state configurations. Clean architecture pays off.