Skip to content
Ilayer

Ilayer research

AI that reviews code the way an adversary would

A tool that cries wolf costs more time than it saves. We built a system of several AI agents. It audits smart contracts, the code that moves money on a blockchain. It attacks its own findings, and shows a person only what survives. It has found real vulnerabilities on live protocols.

Seven thin machined gates in a row on a dark optical rail, receding to the right. A violet beam enters the nearest aperture and passes through all seven, ending in a bright point past the last gate.

The situation

  • Automated audit tools bury a reviewer in false positives.
  • One model is not enough. Different models miss different bugs.

What we built

  1. Several models run against the same target and cross-check each other's findings.
  2. Independent agents attack each finding, trying to disprove it. Only survivors ship.
Five glass rings in a row, receding and shrinking to the right, one violet beam passing through every center, flaring brighter at each ring and ending in a small point of light.
A finding passes every gate, or a human never sees it. Illustration, not a screenshot.

Measured

What this system holds

7-gateEvery finding filtered before a human sees itSeven checks every candidate finding has to clear before a person sees it.Our own system, not a client's. The checks filter findings out; they do not prove one is exploitable.
LiveReal bugs on live protocolsHigh and medium findings on audited DeFi protocols.Protocol names withheld by choice. Severity labels are the platforms' own.
709Commits, and it is still being built709 commits into the pipeline behind these findings, and the seven gates are part of that count.Our own repository. It measures how much has been built, not how much it found.
A dark oscilloscope in a machined case with two round knobs and a vent below the screen. One violet trace wanders across the screen from edge to edge.
One reference line, and the system measured against it at every release.

What held

  • We attack our own results the same way on every project we build.

The stack

  • Multi-agent orchestration
  • Rust TUI
  • Next.js
  • SQLite
  • Solidity

Built with

Evals are tests for a large language model's answers. RAG is short for retrieval-augmented generation.

AI agents and answersTesting and reliabilityAll work

Start a project

Ask what this would look like on your records

The system above was one workflow, scoped before it started. Name yours and you get a straight read on whether your data supports it.