Henry Starr

I am an Applied AI / Forward Deployed Engineer at AMD.

Background

AMD

Applied AI / Forward Deployed Engineer · March 2025–present

I built an internal AI workflow tool at AMD that reached 3,000+ users in ten months. I lead four engineers developing it.

My work spans MCP integrations, deterministic workflow setup, and evaluation. I built OpenTelemetry cost attribution to connect token use to the applications generating it, and work directly with engineering teams to deploy and debug their workflows.

The tool saves engineering teams an estimated 1,600+ hours a month by automating manual setup.

Professional background

University of Maryland

I earned a bachelor’s degree in Computer Science with a concentration in Philosophy.

Projects

TypeScript Next.js SQLite / AWS EC2

Aevron

I built a multi-tenant platform for AI chat, lead capture, and Google Calendar booking. Availability checks and duplicate-booking safeguards handle the scheduling flow; conversation analytics help teams review what happens along the way.

Visit Aevron Product preview

Aevron Telemetry product preview showing sample conversation and lead analytics. All numbers are sample data. — Aevron Telemetry · public product preview with sample data.

JavaScript Node.js

JBrancher

JBrancher routes a step through deterministic rules or a bounded set of candidates, then falls back to the existing actor when a decision is uncertain. I built the routing runtime and local workflow learning, with integrations for Claude Code, Codex, and Pi.

I also built a Markdown and skill compactor. Before accepting a smaller candidate, its CI gate checks protected content and scripts, embedding similarity, and a team’s own behavioral tests.

Read the code Compaction & verification Benchmarks

Other repositories

deadweight: A Python CLI that keeps failed approaches in Git, with SQLite search and agent session hooks.
mandiproxy: A local API proxy that translates LLM messages through Mandarin and reports token-use metrics.