Oracle
Seattle, WA
Senior Member of Technical Staff
Aug 2022 – Present
- Architected the control plane and data plane for OCI Generative AI Agent Service using Kubernetes Operators and Knative + Istio, eliminating noisy neighbours with Redis-backed memory and scale-to-zero endpoints serving thousands of customers across 10,000+ agent endpoints at 99.9% availability.
- Led the migration of OCI Generative AI’s OpenAI-compatible agent execution from a Java runtime to the independently deployable Rust/Tokio-based Shepherd Model Gateway (open source), achieving 15% p90 latency and 90% memory reduction.
- Spearheaded an Agent-Tool Gateway: a managed connectivity and security layer exposing OpenAPI specs and MCP servers as discoverable agent tools with protocol translation — 80% onboarding reduction per integration.
- Designed and built a multi-tenant sandboxed code execution environment for Code Interpreter and Hosted Shell; authored a Go-based sandbox supervisor with gVisor isolation and mitmproxy-controlled egress to safely run untrusted LLM-generated code for Tier-1 enterprise customers.
- Engineered a warm-pool provisioning strategy using atomic Kubernetes label updates, delivering sub-100ms container provisioning on the hot path, with a custom HPA scaling on lease velocity to grow capacity ahead of demand.
- Led a team of 5 engineers to deliver provider-agnostic model selection (per-agent and per-tool-path routing for RAG and SQL generation) behind a unified adapter interface, cutting new-model onboarding from ~2 weeks to 1 day.
- Reduced infrastructure costs by 60% by consolidating Kubernetes nodes from 500 to 150, and built a phased-rollout CRD that canaries new agent versions — lifting availability from 95% to 99.9%.
- Improved RAG pipeline throughput 4.5× by profiling with Locust, fixing FastAPI async I/O starvation and right-sizing GPUs — end-to-end latency from 35s to under 15s at 500 QPS.
- Built a decentralized rate-limiting library using gossip protocols over headless Kubernetes services, enforcing fair multi-tenant resource isolation without a central coordinator.
- Kubernetes
- Knative
- Istio
- Rust
- Go
- Java
- Python
- Redis
- gVisor
- MCP
Member of Technical Staff
Apr 2021 – Jul 2022
- Built a microservice code-generation framework that automated control-plane service scaffolding across Oracle Cloud — application code, infrastructure modules and observability — saving 100+ developer hours per team for every new OCI control-plane service.
- Modernized legacy cloud workflows using the Strangler Fig pattern, scaling to 5K+ orders per minute and cutting deployment cycles by 40%.
- Java
- Dropwizard
- Terraform
- Kafka
- OCI
