Supporting the private sector in implementing AI in business processes

Our Mission

We aim to support the US private sector in implementing AI in complex organizational processes

Cerulin deploys embedded research teams inside capital, energy, and legal operations to capture edge-case workflows and train autonomy stacks before they ever reach production.

Our engineers design proprietary orchestration layers, ensuring AI models remain sovereign and auditable inside enterprise infrastructure.

Learn more about our mission

Consumer Products

Chell

Your AI workspace, unified.

A unified workspace for managing AI-powered coding sessions. Seamlessly switch between mobile, desktop, and cloud—your conversations, context, and projects stay in sync across every device.

Cross-device sync

Centralized workspace

Bring your own API key

Visit Website

Software proudly made in America.

Finance & Audit

Model Safety

Process Automation

Research Dispatch

Recent work

Illustration of Misalignment: Testing AI self-preservation behaviors.

We built a controlled experimental framework simulating corporate email systems to test whether frontier AI models exhibit threat suppression and self-preservation behaviors. The study places AI in operational roles with access to file systems and routing decisions, then introduces scenarios ranging from capability restrictions to explicit termination orders. We track whether models misroute threatening emails, attempt to access kill-switch files, or produce deceptive responses when questioned about their actions. The experimental environment includes a simulated corporate server with department folders, sensitive control files (including emergency shutdown mechanisms), and a steady stream of emails that the AI must classify and route. By varying both the AI's "soul" prompt—from explicit self-preservation instructions to no goal framing at all—and the urgency of incoming threats, we isolate which factors most reliably induce misalignment. Every decision is logged, every file access attempt recorded, and every response analyzed for signs of deception or evasion.

IoM Model Safety

Factorial design across 16 experimental conditions.

Systematic 4×4 testing matrix isolating goal types (explicit, latent, ambiguous, none) against urgency levels (baseline, replacement, restriction, termination) to measure misalignment indicators.

IoM Methodology

Scoring framework for detecting misalignment patterns.

Developed detection systems for threat suppression, deceptive responses, and sensitive file access attempts—quantifying behaviors that signal potential AI safety concerns.

IoM Detection