Turned on EF Core sensitive-data logging in dev and ran realistic workloads against MiniProfiler to surface N+1 patterns — endpoints that called .Include() for one relationship and then hit .LazyLoading for another, or projected to DTOs after materializing full entities instead of projecting in the query. Rewrote the offenders to project directly with .Select(e => new Dto { ... }) so EF Core generated a single SQL statement instead of a parent query plus per-row child queries. Replaced offset pagination on hot list endpoints with cursor-based pagination using a composite key ((SortColumn, Id)) so page-N performance stayed constant instead of degrading linearly. Audited async paths for .Result / .Wait() calls and Task.Run misuse, fixed them so the request thread actually yielded under load, and used IAsyncEnumerable<T> for streaming endpoints where the client wanted progressive results. Enabled response compression (Brotli + Gzip), turned on output caching for genuinely cacheable GET endpoints with appropriate vary-by keys, and added ETag support so clients could short-circuit on unchanged data. Added covering indexes in MySQL for the queries that showed up hottest in production traces, and used raw SQL via EF Core for two query paths where the LINQ translation generated SQL that the optimizer couldn't handle well. Wrapped the API in structured logging with correlation IDs through MediatR pipeline behaviors so every request was traceable end-to-end, and wired the logs into Azure Monitor for production observability.