Kazusa Tsubota

Experience

1. ScalyX.ai

   AWS AI Infrastructure / AI Production Engineer

   Jan 2025 – Present

   USA · Remote

   • Architected and migrated a multi-tenant AI-powered retail platform from per-tenant EC2 to a shared Kubernetes platform on AWS EKS, consolidating workloads across 30+ tenants and enabling demand-based autoscaling while reducing average monthly infrastructure costs by approximately 17%.
   • Designed and optimized vLLM-based GPU inference infrastructure on Kubernetes/EKS, evaluating model characteristics, GPU requirements, workload patterns, and serving configurations to balance inference latency, throughput, availability, and cost.
   • Implemented an AI Gateway and multi-model routing layer across LLM, speech, and vision workloads, routing requests based on workload and resource requirements and reducing GPU-related infrastructure costs by 15% while maintaining target service performance.
   • Built production AI services and workflows using Python/FastAPI, PostgreSQL, Docker, embeddings, semantic search, pgvector, RAG, and LLM tool calling, connecting retrieval, generation, and controlled backend actions.
   • Designed scalable Kubernetes workloads using ALB, health checks, HPA, cluster/node autoscaling, and Helm, while standardizing infrastructure deployment with Terraform, Docker, ECR, and CI/CD.
   • Built end-to-end observability with Prometheus, Grafana, OpenTelemetry, Jaeger, and CloudWatch, reducing average incident investigation time from 50 minutes to 30 minutes (40% reduction).
   • Applied AWS/Kubernetes security and tenant-isolation controls using IAM, VPC/private networking, Security Groups, Kubernetes RBAC, Secrets Manager, and tenant-level access controls; performed synthetic load testing that validated aggregate request throughput of up to 10K requests/sec under controlled load while evaluating latency, concurrency, and GPU utilization.

2. Madoromi, Inc.

   Infrastructure / Backend Engineer

   Jan 2022 – Dec 2024

   Tokyo, Japan · Hybrid

   • Consolidated six backend services from six dedicated EC2 instances into a shared t3.medium environment using NestJS monorepo and Turborepo, preserving logical service boundaries while reducing the production EC2 footprint by 83% (6→1) and monthly EC2 cost from ~$45 to ~$30 (33%) for Alpha/MVP traffic.
   • Designed and operated AWS production infrastructure across EC2 and Aurora RDS, right-sizing compute and database capacity against observed workload patterns rather than maintaining independent per-service capacity.
   • Established a modular TypeScript/NestJS backend architecture with PostgreSQL and shared domain/API boundaries, collaborating with frontend engineers on feature specifications, API contracts, and implementation requirements while structuring services for future containerization and Kubernetes/EKS adoption.
   • Automated non-production infrastructure scheduling with AWS Lambda and EventBridge, reducing development-environment runtime by approximately 70% (~50 vs. ~168 hours/week per instance) while preserving availability during engineering hours.

3. StoreHub

   Full-Stack Engineer

   Dec 2018 – Dec 2021

   Malaysia · Remote

   • Built commerce and POS capabilities using Next.js, TypeScript, React, Node.js, REST APIs, and PostgreSQL, supporting transaction, catalog, inventory, customer, and multi-location merchant workflows across a cloud SaaS platform.
   • Designed reliable distributed and asynchronous systems using idempotent APIs, Redis Streams consumer groups, background processing, and business-rule validation for e-commerce, POS, inventory synchronization, order processing, and notifications.
   • Optimized high-traffic product and inventory services with tenant-scoped Redis caching and PostgreSQL query/index/schema optimization, achieving a 90%+ cache hit rate and processing 1,000+ notifications/sec with <150ms p95 delivery latency.