Paul Serban
6+ years turning complex architectural constraints into production systems. I build agentic AI control planes, high-throughput platforms, and the internal tooling that removes bottlenecks - eliminating process cycle overhead and cutting operations from hours to minutes. The engineer you hire when you need systems that scale and architecture decisions that age well.
Career
Apr 2024 - present
Architected and led an AI-powered testing platform (Python - AWS Bedrock - AgentCore) that cut QA for manual repetitive process by ~90%. Owned 3rd-party game-provider integration across MERN microservices on K8s/AWS.
Nov 2019 - Apr 2024
Built enterprise front-ends on Adobe Experience Manager for financial, automotive and healthcare clients. Cut build time 50% and bundle size 30% (Webpack -> Vite), reduced new-developer onboarding from 3 weeks to 1 via ITCSS/BEM, and delivered WCAG AA across 26 Swiss locales on a 1.5M-page banking site.
Jan 2019 - Nov 2019
Full-service freelance practice for EU clients across MERN, Python, AWS and DevOps.
Toolkit
Not a keyword wall - the tools I reach for and the value each one delivers.
AI & Backend
Python
AI/agent workflows on AWS Bedrock & AgentCore - turning LLMs into production test automation that cut QA ~90%.
Node.js
Event-driven microservices & internal tooling - automated a 6-hour daily bottleneck down to 6 minutes.
TypeScript
End-to-end type safety across MERN monorepos - fewer production incidents, safer refactors.
Cloud & Platform
AWS
Bedrock, S3/CloudFront, Lambda, EC2, IAM - cost-optimised, secure delivery at scale.
Kubernetes
Orchestrates production microservices - zero-downtime, multi-environment releases.
Docker
Reproducible environments - consistent parity from local to production.
Terraform
Infrastructure as code - auditable, repeatable platform provisioning.
Frontend
React
Component architecture for high-traffic platforms - accessible, performant UI.
Next.js
SSR/SSG migrations that lift Core Web Vitals - cut frustrated clicks 75%.
SCSS
ITCSS/BEM design systems - onboarding cut from 3 weeks to 1, zero style regressions.
Observability & CI/CD
Datadog
Latency & error dashboards per provider - surfaced fixes that lifted retention 50%.
Grafana
Grafana/Loki/Prometheus stacks - full-stack monitoring without vendor lock-in.
GitHub Actions
Multi-env pipelines with promotion & rollback - releases are push-button.
Selected work
Full working system — not yet load-tested
Problem. Portfolio, blog, news digest, and quiz each needed the same material - content lived in multiple places, drifted apart, and publish was a manual ritual.
Outcome. Publishing collapsed to one pipeline command; new content types are a migration plus an ingest mapping; CI ingests once and builds four surfaces in parallel.
Full working system — not yet load-tested
Problem. Console-proven AWS hosting for four production domains still needed repeatable provisioning, environment isolation, and a release path that could promote the same four surfaces without long-lived AWS keys or rebuild-per-host chaos.
Outcome. Every non-local environment is provisioned the same way and deployed the same way - fixtures for automation, live content for release rehearsal, public prod behind approval - with zero AWS access keys in CI.
Full working system — not yet load-tested
Problem. Independently deployed static surfaces had no honest telemetry story - the AWS console is not observability, cookie analytics is the wrong privacy default, a static site has nothing for Prometheus to scrape, and "is the bill still cheap?" lived in another console.
Outcome. Phase 0 is live on the laptop - four Grafana dashboards (CloudFront health, traffic, AWS billing, Cloudflare Web Analytics). Phase 0b beacons are in production and correlated against CloudWatch in the same pane. Remaining VPS phases stay gated.
Full working service — not yet load-tested
Problem. Studying from personal notes meant a backend, an account, or a paid SaaS - with no algorithm swap and no reuse of the blog's content.
Outcome. Offline-capable study tool that reuses blog content with zero duplication, lossless algorithm migration, and a harness that compares review load at the same retention target.
Full working system — not yet load-tested
Problem. Four independently deployable frontends needed production HTTPS subdomains with CDN delivery, private origins, and deep-link behaviour that matched how the apps actually emit HTML - without waiting on IaC or CI.
Outcome. Live production URLs for all four surfaces with TLS, edge caching, and correct deep links; a proven reference topology later encoded in Terraform modules.
Full working system — not yet load-tested
Problem. Four monorepo frontends each wanted their own dev port; cross-site navigation broke, and nothing matched the real production domain topology.
Outcome. One `docker compose up` boots the full system at local.*.paulserban.eu on ports 80/443.
Featured writing
July 24, 2026
Learn how to architect a production-grade MCP server: protocol internals, transports, tool design, security, and real TypeScript and Python implementation patterns.
Read post →More reading