Achievements & Certifications

A proof of concept built on GitHub Agentic Workflows (gh-aw), where a team of GitHub Copilot agents running inside GitHub Actions takes a user story from any work tracker, such as Jira or Azure DevOps, through an API, CLI or MCP server, all the way through the delivery lifecycle: plan, code, tests, independent review on a different model, merge, release through the existing pipeline, and the story closed in the tracker. People approve the plan against the exact commit, the merge, and each environment. Identity is passwordless, with GitHub OIDC and federated credentials, and every step is a GitHub Actions job whose logs can be streamed to Azure Monitor, Splunk or Datadog.

A proof of concept built on my own time, to see how a retrieval platform behaves end to end before relying on one. .NET 10 and Microsoft Agent Framework, PostgreSQL pgvector with vector and full-text search merged by Reciprocal Rank Fusion, cited answers, and a LiteLLM gateway so the model provider is a configuration choice. It runs in Docker Compose. AKS and KEDA manifests are drafted, and the gaps before a real deployment are written down.

Traced an unexplained Log Analytics cost leak to its root causes and eliminated it, cutting log-ingestion volume by about 90% through telemetry standards and cost awareness and producing substantial recurring cloud savings.

Led the migration from end-of-life NGINX ingress to Istio Gateway API and a sidecarless ambient mesh across production workloads, ahead of the November 2026 support cutoff, without interrupting the business. The broader zero-trust and identity work covers end-to-end TLS through App Gateway and AKS with cert-manager and Key Vault, private Microsoft Fabric connectivity, customer and partner identity (Entra ID, Entra B2B, Azure AD B2C / Entra External ID, passwordless authentication with Microsoft Authenticator), federated workload credentials with no stored secrets, and privileged access management with CyberArk PAM and Entra PIM.

An Architecture Review Board records decisions as ADRs, with the trade-offs written down. An Enterprise Standards Domain and playbooks use RFC 2119 requirement levels (MUST / SHOULD / MAY), adopted enterprise-wide starting with Infrastructure and DevOps, and keep repository, configuration, deployment, and code ownership separate so teams know who decides what.

Launched a shared Agent Plugins Marketplace for GitHub Copilot and Cursor, built an AI pull-request reviewer that cites the exact standard behind each finding and ranks it by severity, and proved a repeatable pattern for exposing enterprise APIs to AI agents through MCP under Entra ID and API Management. Also built an AI-agent document splitting and classification solution to replace retired ABBYY FlexiCapture, modeled at roughly 10% of the legacy cost.

Enterprise architecture direction across the IT portfolio, 30+ concurrent projects and initiatives from early outlines and intake through delivery, supporting 8+ delivery teams on Azure. Production in West US 3 is designed to run across three availability zones. Lead architect for five strategic initiatives — four mortgage-technology vendor integrations and a document-management cloud migration — and built a shared GitHub Actions CI/CD orchestrator from concept to production.

A multimodal nutrition tool on Gemini: a React web app, a Telegram bot, and a containerized backend on Azure Container Apps, with CI/CD. A personal project, done alongside full-time work.

A similar platform for wine: label recognition, search grounding, and visual taste profiles. The NutriScan architecture was reused, so this one took much less time.

With the engineering teams, 25+ enterprise applications moved to AWS. They were redesigned as they moved, rather than copied as they were. Infrastructure cost came down by about 25%.

A monolith was split into microservices while delivery continued. Deployments became more frequent, and outages tended to be smaller.

Web and desktop tools for a set-top-box testing system. The cost of that system fell by about 80%, and deployments that had taken several days could usually be finished the same day.