Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Why I gave up on agent voting: hard escalation in a 3-reviewer LLM pipeline
Learn how a 3-reviewer LLM pipeline was rebuilt with hard escalation to prevent critical safety findings from being outvoted, ensuring reliable verdicts.
Redline is a production-readiness review tool for AI agents. Three specialized LLM reviewers (Engineer, Risk, Business) read an agent transcript against a company’s pasted rulebook, mark up the dossier in three colors, and a synthesizer returns a Deploy, Hold, or Kill verdict.
Live demo: pasting a real agent dossier into sanjitkangovi.com/redline, watching the three reviewers annotate in parallel, then walking through the deterministic synthesis layer in the backend that decides the final verdict. I’ll show the prompt structure for each reviewer role, the JSON contract between reviewer and synthesizer, and the server-side escalation logic firing when a critical Risk finding contradicts the other two reviewers.