At 3 a.m. an API returned a 500, and it took the on‑call engineer forty‑five minutes to decipher the problem. The delay wasn’t the fault of a complex bug but of an overengineered error handling stack that flooded logs with dozens of fields, stack traces and custom codes. The noise masked the real failure.

The team stripped the system down to three categories—client, server and dependency failures—discarding the 50‑plus custom exception classes. Logs were reformatted into single‑line, grep‑friendly entries like “/users/profile – database connection failed after 3 retries”. Each line also suggested a concrete next step, turning errors into actionable guidance.

Six months later the on‑call rotation resolved incidents in minutes, not hours, and developers voluntarily adopted the lean pattern. The team added lightweight AI helpers such as Crompt AI and Claude Sonnet to surface repeat offenders, but the core win remained a human‑centric log format. Rohit advises testing logs by reading them, not just by unit tests.