Project summary

We are building what we call the Immune System for AI.

Most AI safety and alignment work focuses on training better models. That approach assumes that if a model is aligned at training time, it will behave correctly after deployment. In practice, that assumption fails.

Once models are live, they encounter new inputs, shifting contexts, and evolving user behavior. They begin to hallucinate, drift, or produce outputs that are misleading but still plausible. These failures are not rare edge cases. They are structural.

This project focuses on what happens after deployment.

We are building a system that sits on top of existing models and continuously monitors their behavior, detects when outputs become unreliable or deceptive, and intervenes before those outputs propagate to users or downstream systems.

The goal is simple: make AI systems that can be trusted in real-world conditions, not just in evaluation environments.

What are this project's goals? How will you achieve them?

The goal is to move from static alignment to continuous alignment.

We are building three core capabilities:

1. Real-time detection of unreliable outputs
Identify hallucinations, overconfident claims, and misleading responses as they occur. This includes detecting when outputs appear plausible but lack grounding or internal consistency.

2. Diagnosis of model behavior
Use lightweight interpretability signals and behavioral analysis to understand why a model produced a given output. This is not full transparency, but enough to determine whether the output should be trusted.

3. Intervention and enforcement
Act on failures in real time. This includes flagging, blocking, constraining, or rewriting outputs, as well as applying targeted steering to keep behavior within defined bounds.

We will test and refine this system in real-world or production-like environments where incorrect outputs have consequences. Success is measured by reducing misleading outputs and improving trust calibration, not just benchmark scores.

How will this funding be used?

Funding will be used to accelerate development of the core system:

• Engineering the real-time monitoring and detection layer

• Developing and testing intervention mechanisms

• Running evaluations in production-like environments

• Building datasets and benchmarks based on real-world failure cases

• Publishing findings and open frameworks where appropriate

The focus is on turning this into a working system, not just a research concept.

Who is on your team? What's your track record on similar projects?

The project is led by Owen Sakawa, Founder and CEO of Elloe AI.

Owen has experience building AI systems and infrastructure focused on real-world deployment, particularly in environments where correctness and reliability matter. His current work at Elloe AI is centered on AI safety, compliance, and governance, with a focus on detecting and correcting model behavior in production systems.

This project builds directly on that work. It is not a new direction, but an extension of systems already being designed and tested.

What are the most likely causes and outcomes if this project fails?

The main risks are:

• Detection accuracy is not strong enough to reliably distinguish misleading outputs from acceptable variance

• Interventions degrade model performance or user experience

• The system introduces latency or complexity that limits adoption

If the project fails, the likely outcome is that real-time enforcement proves harder than expected, and alignment remains primarily a training-time problem.

Even in that case, the work will surface important insights about how models fail in production and where current alignment approaches break down.

How much money have you raised in the last 12 months, and from where?

This work has not been directly funded as a standalone research project in the last 12 months.