Kairos is an open-source fraud and risk engine, built for AI agents: Reveneau

Why it mattersA fintech team that pays Sardine or Sift can run the same decision loop on its own servers. An AI coding agent can define new rules through MCP.

Finic AI released Kairos on GitHub under Apache 2.0. The company put the project up on Hacker News on Monday as a Show HN. The thread reached 73 points on 12 comments. The repo had 110 stars five days after creation.

The project is a fraud and risk engine for fintech. It decides whether to approve, step up, review or block each sign-up, login, order, payment and payout. Finic's own page names Sardine, Sift, Riskified and SEON as the paid tools it stands against.

What it does

Kairos runs two kinds of decision model side by side. The first is standard machine learning: logistic regression and XGBoost. The second is Finic's own model, called Grayson. Grayson works with few labels and reads unstructured data. That is a common shape for early-stage fintechs that have not yet built a labelled dataset. The repo says the two run together so a team does not have to pick.

Each decision comes with a risk score and a reason. A review queue and case management are built in. The project supports Shopify for e-commerce fraud with no extra setup.

The agent-first part

Kairos exposes an MCP server. A coding agent can call it to build a workflow, design a model experiment and promote the winning variant. Finic positions this as a way to match the speed at which fraudsters change tactics.

The founder, Ayan, posted the launch and answered questions on the Hacker News thread. One veteran from the AOL spam team in 1995 pushed back on the premise. Fraudsters, he wrote, have always worked around new defences in a single weekend. Another commenter flagged that "Kairos" is already the name of a Linux distro, an identity firm, and a DARPA program. Finic did not change the name.

The stack

The backend is written in Hono on Node.js with Postgres and Drizzle ORM. The console is React and Vite. Finic provides a TypeScript SDK (@finic-ai/kairos) and a Python SDK (finic-kairos). It runs through Docker on a team's own servers. A hosted version lives at kairos.finic.ai.

What this means for a fintech risk team

A risk team already paying for Sardine, Sift or Riskified can run Kairos on shadow traffic. That is the quickest way to see whether the new engine blocks the same bad traffic as the old one. Live decisions do not have to change during the test. The MCP server is the part that is hard to copy from the paid tools. None of them are built to be driven by a coding agent.

The caveats are real. A product on day five with 110 stars has not been stress-tested at the scale of a bank. No named fintech has posted that it runs Kairos in production. Grayson's score on a public fraud benchmark is not published. A team planning a swap should run it against its own historical data for at least a month.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.