Industry
I build production infrastructure and applications for agentic AI systems—focusing on reliability, latency, tenant isolation, and human-governed autonomy in real operational environments.
Ventures & Deployments
Building a multi-tenant AgentOS and production agent infrastructure. Focused on governed autonomous execution, multi-agent orchestration, context and memory management, and human-in-the-loop review workflows for enterprise operations.
- →Multi-agent orchestration and governed execution boundaries
- →Enterprise knowledge indexing, retrieval, and citation provenance
- →Observability, audit trails, and safety guardrails for autonomous workflows
Designing and deploying high-throughput conversational and voice AI systems under strict latency, accuracy, and compliance requirements.
- →Real-time voice agents and clinical/enterprise support workflows for KeyReply
- →Large-scale multilingual customer support intent classification and dialogue systems at Shopee
- →AI guardrails, evaluation pipelines, and compliance infrastructure
Engineering Principles
Problem-driven architecture. Engineering decisions follow from operational constraints—latency budgets, cost ceilings, and failure modes—rather than hype cycles.
Governed autonomy by default. Agents in production are delegated workers, not superusers. Reliable systems require explicit authorization boundaries, audit logs, and graceful human escalation.
Production-first validation. Benchmarks are a starting point; real-world traffic is the test. I iterate on evaluation harnesses built directly from production failure traces.
Open Source & Code
Selected open-source tools, research implementations, and teaching repositories are available on GitHub (@junhua).