Perle Labs Ceo Ahmed Rashad On Why Ai Needs Verifiable Data Infrastructure

What Happened

Among the startups targeting that problem is Perle Labs, which argues that AI systems require a verifiable chain of custody for their training data, particularly in regulated and high-risk environments. With a focus on building an auditable, credentialed data infrastructure for institutions, Perle has raised $17.5 million to date, with its latest funding round led by Framework Ventures. Other investors include CoinFund, Protagonist, HashKey, and Peer VC. The company reports more than one million annotators contributing over a billion scored data points on its platform.

BeInCrypto: You were part of Scale AI during its hypergrowth phase, including major defense contracts and the Meta investment. What did that experience teach you about where traditional AI data pipelines break?

Market Context

The second thing I learned is that the human element is almost always treated as a cost to be minimized rather than a capability to be developed. The transactional model: pay per task then optimize for throughput just degrades quality over time. It burns through the best contributors. The people who can give you genuinely high-quality, expert-level annotations are not the same people who will sit through a gamified micro-task system for pennies. You have to build differently if you want that caliber of input.

We’ve crossed a billion points scored across our annotator network. That’s not just a volume number, it’s a billion traceable, attributed data contributions from verified humans. That’s the foundation of trustworthy AI training data, and it’s structurally impossible to replicate with anonymous crowd labor.”

Why It Matters

BeInCrypto spoke with Ahmed Rashad, CEO of Perle Labs, on the sidelines of ETHDenver 2026. Rashad previously held an operational leadership role at Scale AI during its hypergrowth phase. In the conversation, he discussed data provenance, model collapse, adversarial risks and why he believes sovereign intelligence will become a prerequisite for deploying AI in critical systems.

The second meaning is independence. Acting without outside interference. This is exactly what institutions like the DoD, or an enterprise require when they’re deploying AI in sensitive environments. You cannot have your critical AI infrastructure dependent on data pipelines you don’t control, can’t verify, and can’t defend against tampering. That’s not a theoretical risk. NSA and CISA have both issued operational guidance on data supply chain vulnerabilities as a national security issue.

Compare that to anonymous crowd labor, where a person is essentially fungible. They have no stake in quality because their reputation doesn’t exist, each task is disconnected from the last. The incentive structure produces exactly what you’d expect: minimum viable effort.

Details

AI agents dominated ETHDenver 2026, from autonomous finance to on-chain robotics. But as enthusiasm around “agentic economies” builds, a harder question is emerging: can institutions prove what their AI systems were trained on?

BeInCrypto: You describe Perle Labs as the “sovereign intelligence layer for AI.” For readers who are not inside the data infrastructure debate, what does that actually mean in practical terms?

Ahmed Rashad: “The word sovereign is deliberate, and it carries a few layers.

The most literal meaning is control. If you’re a government, a hospital, a defense contractor, or a large enterprise deploying AI in a high-stakes environment, you need to own the intelligence behind that system, not outsource it to a black box you can’t inspect or audit. Sovereign means you know what your AI was trained on, who validated it, and you can prove it. Most of the industry today cannot say that.

The third meaning is accountability. When AI moves from generating content into making decisions, medical, financial, military, someone has to be able to answer: where did the intelligence come from? Who verified it? Is that record permanent? On Perle, our goal is to have every contribution from every expert annotator is recorded on-chain. It can’t be rewritten. That immutability is what makes the word sovereign accurate rather than just aspirational.

In practical terms, we are building a verification and credentialing layer. If a hospital deploys an AI diagnostic system, it should be able to trace each data point in the training set back to a credentialed professional who validated it. That is sovereign intelligence. That’s what we mean.”

Ahmed Rashad: “Scale was an incredible company. I was there during the period when it went from $90M and now it’s $29B, all of that was taking shape, and I had a front-row seat to where the cracks form.

The fundamental problem is that data quality and scale pull in opposite directions. When you’re growing 100x, the pressure is always to move fast: more data, faster annotation, lower cost per label. And the casualties are precision and accountability. You end up with opaque pipelines: you know roughly what went in, you have some quality metrics on what came out, but the middle is a black box. Who validated this? Were they actually qualified? Was the annotation consistent? Those questions become almost impossible to answer at scale with traditional models.

That realization is what Perle is built on. The data problem isn’t solved by throwing more labor at it. It’s solved by treating contributors as professionals, building verifiable credentialing into the system, and making the entire process auditable end to end.”

BeInCrypto: You’ve reached a million annotators and scored over a billion data points. Most data labeling platforms rely on anonymous crowd labor. What’s structurally different about your reputation model?

Ahmed Rashad: “The core difference is that on Perle, your work history is yours, and it’s permanent. When you complete a task, the record of that contribution, the quality tier it hit, how it compared to expert consensus, is written on-chain. It can’t be edited, can’t be deleted, can’t be reassigned. Over time, that becomes a professional credential that compounds.

Our model inverts that. Contributors build verifiable track records. The platform recognizes domain expertise. For example, a radiologist who consistently produces high-quality medical image annotations builds a profile that reflects that. That reputation drives access to higher-value tasks, better compensation, and more meaningful work. It’s a flywheel: quality compounds because the incentives reward it.

BeInCrypto: Model collapse gets discussed a lot in research circles but rarely makes it into mainstream AI conversations. Why do you think that is, and should more people be worried?

Ahmed Rashad: “It doesn’t make mainstream conversations because it’s a slow-moving crisis, not a dramatic one. Model collapse, where AI systems trained increasingly on AI-generated data start to degrade, lose nuance, and compress toward the mean, doesn’t produce a headline event. It produces a gradual erosion of quality that’s easy to miss until it’s severe.

The mechanism is straightforward: the internet is filling up with AI-generated content. Models trained on that content are learning from their own outputs rather than genuine human knowledge and experience. Each generation of training amplifies the distortions of the last. It’s a feedback loop with no natural correction.

Should more people be worried? Yes, particularly in high-stakes domains. When model collapse affects a content recommendation algorithm, you get worse recommendations. When it affects a medical diagnostic model, a legal reasoning system, or a defense intelligence tool, the consequences are categorically different. The margin for degradation disappears.

This is why the human-verified data layer isn’t optional as AI moves into critical infrastructure. You need a continuous source of genuine, diverse human intelligence to train against; not AI outputs laundered through another model. We have over a million annotators representing genuine domain expertise across dozens of fields. That diversity is the antidote to model collapse. You can’t fix it with synthetic data or more compute.”