Staging environment
Free Lesson

Run AI inside your perimeter, and prove it stayed there

Hosted by Thomas Underhill

Tue, Oct 13, 2026

3:00 PM UTC (30 minutes)

Virtual (Zoom)

Free to join

Invite your network

Go deeper with a course

Where the Data Can't Leave: Self-Hosted AI for Regulated Industries
Thomas Underhill
View syllabus

What you'll learn

Size what you can run on hardware you already own

Model capability by memory tier, stated honestly, including where the gap to frontier APIs still matters.

Prove nothing left the machine during inference

Watch a network monitor during a real query. This is the check almost no deployment actually runs.

Turn a blocked project into a list of technical controls

Egress, retention, training on your inputs, subprocessors. Each objection has a specific answer you can name.

Why this topic matters

Two things changed since your organization decided this was impossible. Models small enough to self-host got good, and waiting stopped being free. The alternative to a sanctioned path is not that staff stop using AI, it is that they use consumer tools on their own phones with your data in them. The objection that stopped you was never about AI. It was about one deployment model, which is a solvable engineering problem.

You'll learn from

Thomas Underhill

Product & Eng Leader (AWS, VMware, HashiCorp, NGINX) · Adjunct Professor

I build and operate AI systems in places where you have to prove things: that the data never left, that the model did what you say, that a person approved the action.

Constellation is a self-hosted platform I architect and maintain: authorization on every tool call, pluggable secrets and sandboxing, a tamper-evident audit trail, local inference, sealed and air-gapped by default. My research is on evaluation that cannot be gamed by whoever benefits from the score.

Twenty five years in enterprise software before that, in senior roles at PayPal, Oracle, VMware, HashiCorp, F5 and AWS, running SOC 2, PCI-DSS, FedRAMP and NIST 800-53 as engineering programs, not paperwork. Nine years as adjunct faculty at universities. U.S. patent. J.D. Startups.

Everything I teach, I have had to make work first.

See all products from Thomas

Sign up to join this lesson

By continuing, you agree to Maven's Terms and Privacy Policy.