The problem AEGIS solves

The Problem

Blind transaction signing causes massive fund losses in Web3. On Sui, this is critical because:

• Complex Multi-Step PTBs: Scammers hide wallet-drainers inside transactions that look normal on the surface.
• Blind Signatures: Wallets display raw hex code and technical parameters instead of human-readable outcomes.
• ~400ms Sub-Second Finality: Sui has no mempool waiting room. Once you sign, funds are gone instantly—post-hack alerts and revoke tools are too late.

What It’s Used For

AEGIS is a browser extension and pre-execution security oracle that inspects transactions before the wallet signs:

• Live Deterministic Dry-Run: Computes exact net balance changes and object mutations against live Sui state.
• Plain-English Explanations: Explains what enters/leaves your wallet with zero technical jargon.
• Dual-Model Consensus (Gonka Router): Dispatches parallel inference across two models (DeepSeek-V4-Flash + MiniMax-M2.7) to eliminate hallucinations and compute a calibrated Truth Score.
• Clear Verdicts: Delivers instant Approve, Caution, or Reject recommendations.
• Gasless zkLogin (Enoki): Allows users to onboard with Web2 logins and transact with sponsored gas.

How It Makes Web3 Safer & Easier

• Stops Exploits Pre-Signature: Intercepts threats before execution, not after funds are lost.
• Zero Custody, Zero Admin Keys: Never holds user funds or contract capabilities—no central point of failure.
• Defense-in-Depth: Automatically enforces the strictest conservative verdict if AI models disagree.
• Permanent Audit Proofs: Records cryptographic audit proofs to Walrus Storage and MongoDB.
• 1-Line dApp SDK: Any Sui dApp can integrate aegis.analyze(tx) with one line of code.

Challenges we ran into

Challenges We Ran Into

1. Chrome MV3 Worker Dying During Confirmation (30s+ Timeout)

   • Challenge: Chrome kills background service workers after ~30s. If a user hesitated before clicking "Confirm", the worker was already dead, dropping the action and freezing the dApp.
   • Fix: Implemented bidirectional chrome.runtime.Port connections and an active 8-second heartbeat interval while confirmations are pending, keeping the worker alive and ensuring reliable routing.

2. Gonka Router 502 Bad Gateway Inconsistency

   • Challenge: Decentralized model calls on Gonka Router occasionally returned 502 Bad Gateway errors due to network node drops.

     

   • Fix: Built a resilient fallback pipeline in LangGraph: if Gonka Router fails or times out, the agent automatically fallsback to Google Gemini without crashing the stream.

3. Building a LangGraph AI Agent as Non-AI Students

   • Challenge: Without formal AI backgrounds, orchestrating cyclical state graphs, tool routing, and preventing LLM hallucinations was a steep learning curve.
   • Fix: Constrained the LLM's routing decisions using strict Zod schemas (PlanSchema) over 3 fixed tools, keeping blockchain dry-runs and risk scoring in 100% deterministic TypeScript.