LLM Inference

• Led a team, from design to implementation stage of a ledger balance service with

"Five 9's" of availability and $550M of money movement processed till date

• Introduced self-correction capability that enabled reconciliation of balances

maintained by the ledger service from 2+ weeks to 2 days

• Designed and built a card authorization system that has a P999 response time of 2

seconds

• Optimized PostgreSQL database to reduce memory utilization by 90%

• Created the control plane for a Kafka based event bus that transports 850k events each day

Created a new monitoring framework that facilitated collection and analysis of

relevant machine learning job metrics to improve model performance

• Devised a log analysis framework that dynamically constructed job graphs that

reduced failed job detection time by 75%

• Developed a data completeness monitoring system for a log data Kafka cluster that

ensured a transfer completeness rate of "Five 9's"

Building distributed systems at Salesforce