engineering ai interviews

How AI Interviews Actually Work

A look inside the two-layer guardrail, the LangGraph agent and the rubric engine that power a single Sovrano interview.

Sovrano Team ·

Every Sovrano interview is a coordinated dance between three systems. This post walks through them, end to end.

The three layers

1. Guardrail

Two-layer ModernBERT classifier. Blocks unsafe content, then routes by intent.

2. Agent

LangGraph multi-agent graph — opening, technical depth, closing.

3. Rubric

Async evaluation pipeline scoring the transcript against a domain rubric.

Why MDX for this post?

This file is .mdx, which means we can embed real Astro / HTML components directly inside the markdown. That’s how those three cards above render with Tailwind classes — no images, no plugins, just components.

Use MDX when you want:

  • Interactive or visually rich explainers
  • Reusable components (CTAs, callouts, comparison tables)
  • Embedded demos, charts or code playgrounds

Use plain .md when you just want to write words. Both live in the same collection and follow the same schema.

The interview in one sentence

The guardrail decides if the model should answer, the agent decides what to ask next, and the rubric decides how good the candidate was.

Everything else is plumbing.

Want to know more?

We’ll be publishing deep-dives on each layer. Subscribe to the RSS feed to follow along.