Deploys were a held-breath event only a couple of people knew how to do safely, so releases bunched up and got riskier the longer we waited. I made shipping boring — automated the pipeline, put the guardrails in code, and handed the whole team a button they could trust at 5pm on a Friday.

Added backpressure and at-least-once delivery with reconciliation; row loss under peak load went to zero.

The pipeline lost rows quietly under load and we only found out when a customer's numbers didn't add up, which is the worst possible way to find out. I chased it to the root — backpressure, at-least-once handling, and reconciliation that proved nothing fell on the floor rather than hoping it hadn't.

Introduced Redis and read replicas behind a clean interface; primary CPU dropped 50%.

Costs were climbing faster than usage and nobody owned the number, so every team assumed someone else was watching it. I took the problem, instrumented spend down to the service level, and worked through the biggest offenders one at a time until the trend line bent and stayed bent.

Right-sized clusters and added autoscaling; the monthly bill dropped 25% at the same SLOs.

Batch jobs ran overnight, so by the time anyone saw a failure the window had closed and the day already started behind. I moved the critical paths to streaming, made retries safe to repeat, and turned a nightly gamble into something that just quietly kept up.

Decomposed the monolith behind a strangler proxy into 7 services — no maintenance window needed.