At Oracle, I designed and shipped the AI gateway and multi-agent orchestration behind an enterprise AI assistant with 100K active users. I built topic-specific RAG agents and an MCP-based workflow that reduced average response latency from five seconds to two.
I built evaluation harnesses from golden and production-derived multi-turn datasets to catch routing regressions and agent failures before release. Orchestrating topic-specific agents raised RAG answer accuracy from 56% to 87% on document-based questions.
I also fine-tuned open-source Llama 3.2 3B Instruct weights with Transformers and PEFT/LoRA for intent classification. Routing accuracy rose from about 70% to 90% on production data, while LLM inference cost fell.
Previously, at Fair Isaac Corporation (FICO), I deployed 28 credit models for APAC and built a Kafka/Flink fraud-detection pipeline that increased throughput from approximately 300 to 500 records per second. I also developed a Python/LangGraph workflow to monitor production ML pipelines, and at Micron Technology I developed and tested low-level software for enterprise UFS 3.1 storage architecture.
