Enterprise AI Architect · Agentic Systems · Austin, TX
I build agent systems that have to run inside real enterprise clouds, not just in a demo. When they catch fire, I put it out and try to make sure it stays out.
I'm Rapha, based in Austin. Most of my work happens inside enterprise clouds,
where a good agent demo runs into private networking, firewalls, and compliance.
I've been pulled into projects where the AI worked fine in a sandbox and fell
over the minute it touched the real environment.
What I've learned is that the model is rarely what breaks. It's the mock data
nobody replaced, the identity that doesn't get passed along, the firewall rule
nobody knew about, the "temporary" fix from three sprints ago. I like those
problems. They're solvable if you're patient and you write things down.
I build with Claude Code most days. When I leave a project, I want the fix
living in the pipeline, not in my head.
Patrol Logs
66degrees
Gemini Architect · May 2026 – Present
Found
Agents that had only ever run on mock data, inside a locked-down enterprise cloud. Egress was blocked, and the caller's identity wasn't making it from one agent to the next.
Did
Worked through the hub-and-spoke networking and firewall egress, and designed the path to a legacy enterprise database. Wrote a security audit of a two-agent production system (about 130 files, 35 findings) and fixed the identity bug. Built and handed off a Slack research agent on Cloud Run.
Prevention
Deploys run through the customer's own IaC and CI/CD pipelines, so nothing only works on someone's laptop. For the Slack agent I built a 60-question eval harness with an LLM judge, so a model change gets measured instead of eyeballed.
Odinzen
Co-founder · 2025 – Present
Found
A small research team building everything with Claude Code, where good tools existed but every new session started from scratch.
Did
Advised on the architecture of an MCP server that gives Claude about two dozen CALPHAD and data tools. In October I took over AI engineering and reviewed all 16 repos.
Prevention
Started with the cheapest fix: a shared CLAUDE.md so Claude Code sessions don't start cold. A shared provenance format and one-command reproduce are next.
Retail and financial services teams who wanted generative AI but had to get it past security and governance first.
Did
Designed AI gateway patterns on Apigee (zero-trust, OAuth 2.0, prompt inspection, semantic caching) and Gemini Enterprise architectures grounded in BigQuery. Advised C-level and IT leaders, and ran technical workshops for 150+ developers.
Prevention
Reusable Terraform patterns and reference architectures that teams could run without me.
