Worked on the BerriAI/litellm repository to address database reliability in high-concurrency environments by fixing a PostgreSQL deadlock issue during spend update operations. Introduced a deterministic lock ordering mechanism within Prisma transactions, ensuring that spend dictionaries—such as user, key, team, and organization—were consistently sorted before updates. This approach prevented deadlocks in multi-pod deployments, reducing downtime and maintaining data integrity under load. Leveraged Python for backend development, asynchronous programming, and database management, with a focus on robust testing. The solution improved scalability and reliability of spend updates, directly supporting business continuity and operational stability in distributed service architectures.