About

Michael Golden

Fairfield, IA · Remote

Sixteen years building and leading backend teams. Now fine-tuning my own models, building agentic systems with guardrails, and shipping with AI as a pair.

I've spent about sixteen years building distributed systems and the teams that ship them — backend services and cloud-native platforms on Azure and AWS, mostly in healthcare technology and SaaS, much of it under HIPAA and SOC 2. I've led engineering teams and organizations of up to 17 people through architecture, API design, CI/CD, and the unglamorous work of keeping production stable.

By early 2025, the company had grown to the point where it needed a dedicated CTO. I chose to stay close to the technical work, moving into a Lead Engineer role focused on backend architecture, API design, and delivery for our enterprise client systems.

Since then I've been going deep on machine learning the way I learn best — by building complete systems, documented decision by decision.

The work on this site

PlayIT is a hybrid Steam recommender — a gradient-boosted model predicts what you’d enjoy, a fine-tuned language model explains why. It runs end to end on a home GPU, and every design call, experiment, and dead end is written up. Try it or read the architecture and key decisions.

FleetAlert AI is an agentic fleet-maintenance investigator — Claude reasons over telemetry and a knowledge base, proposes a fix, and then waits for a human to confirm before anything runs. It’s built on AWS (Step Functions, Lambda, SQS, DynamoDB) with a fixed action whitelist. Alerts arrive three ways — the web, an inbound email, or a detector watching live telemetry — every step is traced, and every change to the agent is measured by evals before it ships. It’s been load-tested to 24,000 investigations an hour without losing one. Try it or read the architecture and guardrails.

Around FleetAlert sits a small agentic toolkit, each piece its own repo with its own decision record: an MCP server that exposes the demo to Claude Code and Claude Desktop with safety tiers and human approval through elicitation; an evaluation harness that grades agent trajectories with confidence intervals, a calibrated judge and a CI gate; and a cost and safety governor that sits in front of the Claude API with per-agent budgets, tool allowlists and redaction.

Projects - Architecture

Elsewhere