Prophet

Demo

About

The problem it solves Last year, Decentralized Finance lost millions due to vulnerabilities and logic errors in smart contracts. Security reviews are slow and hard to act on. Teams wait weeks for audits, get long reports, and still have to turn findings into code changes themselves. There is no single path from “is this safe?” to “here is the fix.” You juggle scanners, auditors, and manual patching instead of one pipeline that finds, explains, and fixes. While developers are increasingly using Web2 AI models such as Anthropic or OpenAI to review their contracts, these models operate as a "black box" with little transparency and verifiability with code changes. DeFi cannot scale safely without better tooling. Contracts move real value, and teams need fast, clear, reliable audits and fixes so they can ship with confidence. Prophet is built for that: one pipeline with transparency and efficiency so security becomes part of the flow, not a bottleneck. Use of AI tools and agents Prophet uses AI through a Zero Gravity agent in three places: When you click Run Analysis, the agent sends your contract to 0G for inference. The model acts as a security auditor and returns a structured report: risk score, findings with locations and explanations. That output is parsed and shown in the UI so you can see what’s wrong and where. When Generate Test Cases is clicked, the agent goes through all potential transactions that could put the contract in a vulnerable position and creates test cases to test the contract on through reinforcement learning. When you click Generate Fix, the agent runs a second AI step with your code and the report. The model returns patched Solidity. The agent cleans the response and the UI shows a diff. You review and Accept or Reject. The AI suggests; you decide. So one flow: analyze with 0G, get the report, create test cases, run them in Dockerized Foundry, generate fixes with AI, then you approve or reject. The AI steps work together and the user always has the final say.

How it's made

Challenges we ran into One big hurdle was getting reliable, structured output from the AI. The analyzer and fix generator need JSON and raw Solidity, not markdown or chat. We had to design prompts and response handling so the model returns something we can parse every time (e.g. strict schema, stripping code fences, validating pragma). Another challenge was the fix-review UX. Showing an AI-generated patch in a way that’s clear and stable (diff view, accept/reject, no cursor jumps or accidental overwrites) took several iterations; we simplified to a read-only diff and a single source of truth so the UI doesn’t fight the user. Integrating 0G inference into the agent (calls, error handling, and fitting it into the existing analyze -> report -> fix flow) also required some iteration so the pipeline stays fast and the frontend stays in sync.