Minglong Pan

AI Platform & Multi-Tenant LLM Serving

• Designed and own multi-tenant LLM serving for real-time GenAI translation on AWS Bedrock — automated tenant onboarding, request routing, and per-tenant cost tracking across regions, with faster tenant onboarding and improved serving efficiency at production scale.
• Built the multi-tenant serving and staged-rollout infrastructure used to migrate production traffic from legacy machine translation to GenAI translation on Bedrock — eval-gated dial-up rollouts across many languages and regions.
• Built a persona-routed MCP server framework (Python FastMCP, stdio) on Anthropic's Model Context Protocol — semantic routing of tasks to persona → skills → tools, with team resource packages and per-team JSON config; adopted by peer teams without forking shared infrastructure.
• Built MCP-based internal AI agents for developer-productivity and incident-response workflows.
• Led an enterprise onboarding across European and Japanese marketplaces — large-scale terminology and multilingual configuration, delivered as a staged production change.

Inference Reliability, Safety & Cost Optimization

• Built guardrails on AWS Bedrock for the GenAI translation stack, screening model output before it reaches customers; extended the translation evaluation framework with standard MT-quality metrics (chrF++, COMET) and LLM-as-judge meaning-preservation checks.
• Built a direction-aware SageMaker scaling tool with dry-run and rollback that significantly right-sized GenAI inference capacity during the MT→GenAI cutover, materially cutting compute cost.
• Built staged prompt-rollout controls with a safe kill-switch and per-tenant gating; hardened batch inference with retry/backoff and concurrency limits for reliability under throttling.

Distributed Platform Engineering & Migrations

• Migrated a regional voice channel to Amazon Connect — resolved cross-region TLS and IAM integration challenges and coordinated multiple partner teams to bring customer-service agents online at scale.
• Migrated a core chat service to a container-based architecture on ECS/Fargate; built load- and stress-testing infrastructure for active chat monitoring, throttling, and persistence across regions.
• Led the cross-organization transfer of a suite of chat services between engineering teams — authored knowledge-transfer documentation and ramped the receiving team with no major operational incidents.

Technical Communication

• Authored internal engineering write-ups on AI developer tooling and MCP-based workflows to socialize reusable AI-infrastructure patterns across teams.

• Built signal-processing and seismic-imaging pipelines in Perl / Java / Python — Fourier transform, Tau-P transform, filter design — on petabyte-scale seismic data spanning 120+ km³ subsurface volume; improved image SNR by 5+ dB.
• Developed tree-based classification algorithm in Java and released to production; improved workflow throughput, contributing to major proprietary client engagements.