Introducing Kairos: open-source fraud decisioning, tuned by your coding agent

TL;DR

Kairos is open source under Apache-2.0: github.com/finic-ai/kairos. You can self-host it with Docker or use the hosted version at kairos.finic.ai.1

Fraud models are hard to keep working

Most fraud teams pick between two imperfect options. Rules are easy to read, but they break the moment fraudsters change tactics. Machine-learning models adapt better, but they need months of chargebacks before they are any good, and someone has to keep retraining and re-testing them as fraud moves.

That someone is usually a data scientist, and most teams don't have one to spare. So models drift, rules pile up, and nobody has time to check whether a new signal actually helps.

What Kairos is

Kairos is a platform for real-time fraud and risk decisions. You send it what happens in your product: sign-ups, logins, orders, payouts and chargebacks. When something risky happens, you ask Kairos what to do, and it answers approve, step up, review or block, with a risk score and the reasons behind it.

Around that sit workflows that decide what to check and in what order, a model registry, review queues and cases for your team, and SDKs for TypeScript (@finic-ai/kairos) and Python (finic-kairos). It covers the ground of platforms like Sift and Sardine, except you can read the code and run it yourself.

Two things make it different: your coding agent tunes it, and it pairs decision models with traditional machine learning.

Your coding agent runs the experiments

Kairos has an MCP server, so Claude Code, Cursor or Codex can work in it directly. An admin decides which tools each agent key may use, and every action is recorded under that key's name. From there, the agent can do the work of a fraud data scientist:

1. 1.Read. It looks at recent decisions, the fraud that got through and the signals behind each score.
2. 2.Propose. It writes a hypothesis and a few variants to test: logistic regression on the order's numbers, XGBoost, Grayson's answers, or a combination.
3. 3.Run on the server. Kairos trains each variant on the oldest labeled data, compares them on newer data and tests the best on the newest. Nothing trains on the agent's laptop.
4. 4.Compare. Each variant gets ROC-AUC and PR-AUC, its gain over a baseline with a 95% interval, and precision and recall at your review and block thresholds.
5. 5.Ship. The agent picks a variant, writes down why, and applies it as a draft for review or as the workflow's next version.

Stripe recently described a similar loop for Radar, built on its own in-house ML infrastructure. Kairos makes it available to teams who don't have that infrastructure.2

Splitting by time matters because fraud reports arrive late, often weeks after the order. Kairos only counts a run as labeled once its outcome window has passed, so recent fraud that hasn't been reported yet doesn't train the model as legitimate.

Decision models next to traditional machine learning

Much of the fraud signal hides in text that a traditional model can't read: a throwaway email, a reshipping address, gift cards in the cart. Grayson, Finic's decision model, reads it. You ask it questions in plain English, such as “Is this email throwaway?”, and it answers with a calibrated probability.3

• ✣It works with few labels. Grayson needs none of your outcomes, so it is useful from day one.
• ✣It reads unstructured data. Emails, names, addresses, notes and item titles all count.
• ✣It explains itself. Every answer is to a question someone wrote, so every decision says why. We describe the approach in Question-Bottleneck Scoring.

In Kairos, Grayson's answers become features next to your own fields in a logistic regression or XGBoost model. We've tested that combination on public data for insurance fraud and loan default. Kairos lets you test it on yours.456

People stay in charge

An agent changing a live fraud model needs limits. In Kairos, an admin decides which tool groups each agent key can use. Every experiment records who ran it, the hypothesis, and why a variant was chosen. You can keep agents to drafts that a person reviews in the builder before anything goes live, and an experiment can only be applied once, to the workflow version it ran on.

Get started

• ✣Cloud: sign up at kairos.finic.ai and connect your agent from Settings › MCP.
• ✣Self-host: clone finic-ai/kairos and run it with Docker. Grayson needs an API key.
• ✣SDKs: npm install @finic-ai/kairos or pip install finic-kairos.
• ✣Shopify: for the Kairos app for your store, reach us on Discord or at support@finic.ai.

We'd like to hear what you'd let an agent change on its own, and what should always need a person. Join the conversation on Discord.