Things I’ve built, and teams I’ve built them with.

The problem

Youth and travel baseball generates a flood of raw data: play-by-play logs, box scores, radar readings and video. Coaches and scouts have no time to turn it into insight, and generic chatbots make things up when they try.

What I built

An agentic AI platform that turns raw play-by-play and video into scouting reports, coaching analytics and player development insight. The backend is an agentic workload manager: it breaks each analysis request into steps, routes them across specialized data and reasoning agents, and returns structured, source-grounded output. The analytics layer is exposed through the Model Context Protocol (MCP), so AI assistants can query it directly as a tool.

To choose models with evidence rather than instinct, I built a multi-model evaluation harness that compares LLMs on video extraction and interpretation. The product itself was built with AI coding agents (Kiro, Claude Code and OpenAI Codex) as primary builders, with me as architect and reviewer.

Why it matters to you

This is the same pattern that works in the enterprise: decompose the work, ground every answer in source data, evaluate models objectively and expose capabilities through open protocols.

• Live pilotRunning with a travel baseball organization
• MCPAnalytics available to AI assistants as tools
• Multi-modelEvaluation harness drives model selection

The problem

Tournament staff needed to chart every pitch (type, velocity and location) from a phone at the ballfield, where the cell signal comes and goes, without ever losing or double-counting data.

What I built

A phone-first app that saves every pitch on the device first and syncs in order when the signal returns. Every operation carries a client-generated ID, so retries are idempotent and never double count. A service worker caches the app, schedule and rosters, so a game loaded in the parking lot still opens with no signal.

Production runs serverless on AWS (CloudFront and S3 for the app, an HTTP API on Lambda and DynamoDB for data), defined in AWS CDK. It costs close to nothing between tournaments. Reference data such as tournaments, schedules and rosters is pulled from the Two Out Barrels platform through its MCP interface.

• Offline-firstWorks with no signal and syncs safely later
• Near-zero idle costPay-per-request serverless design
• Director viewsLeaderboards, arsenals and CSV export

The challenge

Before generative AI reached the public market, the largest enterprises in the Americas needed to understand what Azure OpenAI and GitHub Copilot could do, and the product teams needed to learn what customers actually needed.

What we did

I established Microsoft’s earliest field go-to-market for Azure OpenAI and GitHub Copilot and ran more than 100 enterprise engagements that shaped both how customers adopted the technology and the product feedback behind it. I authored the strategic account plan for the region’s top 50 accounts, covering $953M in Azure consumption, and set multi-year application modernization and AI direction for those enterprises.

Through several company reorganizations, the team stayed intact and sustained $650M in annual Azure revenue.

• 100+Early enterprise generative AI engagements
• $953MTop-50 strategic account plan
• 125+Published customer stories

The idea

Training sessions teach features; building something real changes minds. With one of the best engineers I’ve worked with, I created Codeslingers: half a day training a customer’s teams on the practical fundamentals of adopting AI, then two days in a competition-style hackathon building a real solution to a scenario they chose, with their business, their data and their problem. Every team presented, and the executive sponsor chose a winner.

What happened

We ran more than twenty-five of these engagements, and multiple use cases went into production as working systems, not demo apps that died the Monday after. Weeks later the deeper follow-up questions would arrive: where else AI could help, and how the cloud architecture should be shaped.

It worked because of what happened in the room. We showed teams how to adopt the technology and how to work together, and those turned out to be the same problem. It is still one of the fastest ways I know to get a team building with AI.

• 25+Customer engagements
• ProductionMultiple use cases shipped as working systems
• 2.5 daysFrom fundamentals to a presented solution

The challenge

Build an entirely new commercial coverage motion for public sector and education customers, as a strategic company pilot, and reach accounts the field had no efficient way to reach.

What we did

I grew the organization from 12 to 45 people across four countries and sponsored the development of an agent-driven prospecting tool that combined market and firmographic data to open under-developed subsidiary accounts. I also created a technical upskilling program and earned five foundational certifications myself in the first two weeks to set the standard.

• $15M+Net-new white-space revenue from 180+ opportunities
• 100%Team certification; 80% across the organization
• 92Leadership effectiveness score

What we built

The cloud-native teams I led and worked in published open workshops and reference implementations so that customers could adopt new platforms safely and quickly. Highlights include an advanced AKS secure baseline workshop, a webinar series on configuring, securing and extending Azure Kubernetes Service, Azure Arc for Kubernetes GitOps and policy demos, and dozens of serverless, event-driven, Service Fabric, IoT and bot samples.

Earlier, as an application development specialist, I architected cloud-native and modernization solutions across Cognitive Services, Service Fabric and AKS, reaching 256% of quota and earning the GBB Global Impact Award.

• 256%Quota attainment as a technical specialist
• Global ImpactGBB award for customer outcomes
• OpenWorkshops and samples published on GitHub

What I built

fiatnovena.com is an installable, offline-capable prayer app that keeps every user’s data on their own device. Its content is generated from a sourced research pipeline and checked by automated tests. serviamrex.com, the site you are reading, is built the same way.

Both run on a private S3 bucket behind CloudFront with HTTPS, strict security headers, a content security policy and a monthly budget alert. Everything is defined in CloudFormation. GitHub Actions deploys use OIDC for short-lived credentials, so no AWS keys are stored anywhere, and a failing test stops a deploy. This site’s contact form is a small Lambda function sending through Amazon SES.

• ~$1/monthTypical hosting cost at small-business traffic
• No stored keysOIDC-based deployments from GitHub
• TestedEvery deploy runs the test suite first