The problem Aegis solves

Aegis watches on-chain money movements across bridges, exchanges, and escrows for early signs of a hack in progress. When it detects something anomalous, such as a sudden outflow far beyond a venue's normal activity, it does not act on that signal alone. Instead it sends the evidence to two independent AI models running on the decentralized Gonka Network, and only proceeds if both models agree the incident is genuinely real. Once confirmed, Aegis automatically buys a protective put option on Thetanuts Finance, so a user holding ETH or another asset receives a cash payout if the price falls sharply because of the hack, all without ever taking custody of the user's coins. In short, Aegis provides automatic, AI-verified insurance against crypto disasters that is triggered and executed within seconds, while keeping every AI decision publicly auditable rather than hidden inside a black box.

In short: automatic, AI-verified insurance against crypto disasters, triggered and executed in seconds, with every AI decision publicly auditable — not a black box making trades on your behalf.

Challenges we ran into

One major challenge was that Thetanuts' own SDK documentation turned out to be outdated by a full version, which meant we had to manually audit our integration against the actual SDK source code line by line. That audit uncovered around ten incorrect assumptions in our code, and the effect was severe: both of our trading paths were completely broken, with every trade attempt failing silently in the background.

Another challenge came from the unpredictability of the AI models themselves. Asking the same model the exact same question twice, even with randomness turned all the way down, produced wildly different confidence scores ranging from 55 to 82. We addressed this by querying each model five times and using the median result instead of relying on any single answer.

We also learned that simply averaging AI opinions can be actively misleading. If one model reports being 95 percent confident that an event is a hack while another reports only 20 percent confidence, averaging those two numbers produces 57 percent, a figure that does not actually reflect either model's real judgment. To avoid this trap, we require both models to land clearly on the same side of the decision threshold, so that a split verdict simply blocks the trade rather than being smoothed over by an average.

The Gonka AI network itself proved unreliable in ways that were not obvious at first glance. Its model listing sometimes advertised models that were not actually available, and individual models would occasionally rate limit for over 40 minutes before working normally again. To handle this, we built in retry and backoff logic and made sure that "the AI network is unreachable" is always treated as its own distinct state, never confused with "the AI evaluated this and found nothing wrong."

Finally, we had to work through the challenge of calibrating exactly when a confirmed verdict should be trusted. Early in development, the AI significantly under scored a genuine incident modeled on a real 1.4 billion dollar exchange hack. We resolved this by anchoring the AI's scoring against concrete real and fake examples rather than relying on a vague general scale, so the confirmation threshold is now something we measured carefully rather than something we simply guessed.