Client projects first, with names and identifying details left out. After those, three projects of my own, where I can show every number.
Client work
Same traffic, a fifth of the bill. Most of the money was idle capacity, hot storage nobody read and data crossing zones for no reason.
Client work
A thousand times more people, and a platform where one part can change without stopping the rest.
Client work
One supported way to run experiments and release models, in two clouds, without pretending the clouds share storage.
Client work
Tripling request capacity for a launch, and the parts of the transaction autoscaling can't touch.
Client work
A paved road for the common case, with the guardrails built in.
Client work
Permission checks before retrieval, citations on every draft, and a person who decides what gets sent.
Client work
Keep the raw evidence, quarantine what doesn't parse, and make replays idempotent.
My own project
One node, one model, one user. The setup that made it repeatable.
My own project
CRDs for profiles and runs, a scheduler that checks headroom, and a database that remembers every attempt.
My own project
Build where the cache already is, and never run the same validation twice.
If something here sounds like your problem, send me a few lines about what you're running.