Savant Chat

Savant Chat is an advanced AI-driven platform that audits Solidity smart contracts for potential security threats and vulnerabilities. Leveraging the deepseek and voyage models, it automatically pinpoints critical and major issues—such as reentrancy attacks, overflow/underflow bugs, and access control flaws—before they can be exploited. Savant Chat empowers developers with rapid, accurate, and actionable insights, helping them build and maintain secure blockchain solutions without the complexity of manual code reviews.

Savant Chat operates as an AI agent that orchestrates multiple steps in a layered auditing process. First, we use tree-sitter and voyage-code-3 to parse and segment Solidity code, forming a detailed context for the model to work with. Then, for each segment, our AI agent adopts a “paranoid” stance—evaluating which classes of vulnerabilities could theoretically exist in that snippet.

Next, we leverage a reasoning LLM, running it multiple times per segment under the assumption that a specific class of vulnerability is definitely present, prompting the LLM to localize it precisely. Any identified vulnerabilities are then forwarded to a dedicated critic model, which filters out false positives to ensure higher accuracy.

Beyond this core flow, we’re experimenting with further techniques to enhance deep analysis. These include selectively rewriting and simplifying code to better expose security issues and exploring the execution of code in virtual machines for dynamic testing.