Finic open-sources Kairos, an agent-tuned fraud and risk decisioning platform

Finic has open-sourced Kairos, a real-time risk and fraud decisioning platform pitched as an alternative to commercial services like Sardine, Sift, Riskified and SEON. The launch appeared on the front page of Hacker News as a Show HN post linking to the project's GitHub repository, finic-ai/kairos. The two differentiators the project itself emphasizes: coding agents do the tuning work, and its "decision model" approach is meant to catch fraud before a company has accumulated enough labeled outcomes to train a conventional model.

What it decides, and how

Kairos evaluates events as they happen — sign-ups, logins, orders, payments and payouts — and returns one of four verdicts: approve, step up, review, or block. Each verdict arrives with a risk score and the reasons behind it, drawn from the subject's history, devices, cards and addresses shared with other subjects, and session data. Every decision is stored alongside the signals that produced it.

Decisions are expressed as workflows: ask a decision model a question, score with a machine learning model, branch on the result, call webhooks, open cases, or defer a final call to later. Cases land in review queues for people, and those human resolutions become the labels used to train future models.

Agents as the fraud ML engineer

The unusual part is who tunes the system. Kairos ships an MCP server, so agents such as Claude Code, Cursor, Codex or VS Code-based assistants can connect with an API key, read past decisions, and build workflows themselves, with every agent action recorded under its own key.

Agents can also run experiments. The agent proposes model variants — a decision model alone, logistic regression on numeric fields, gradient-boosted trees, or a stacked combination — and Kairos trains each one, selecting on an intermediate slice of labels and testing on the most recent. It then reports ROC-AUC, PR-AUC, precision and recall at chosen thresholds, and the gain over a baseline with a 95 percent interval. The agent explains its pick, publishes the winning variant as the workflow's next version, and returns to re-evaluate when new labels arrive.

Two model families, combined

The platform pairs two approaches. Grayson, Finic's decision model, is asked plain-language questions about each request — for example, whether a shipping address resembles a reshipper, or whether an email address looks disposable — and answers with calibrated probabilities. According to the repository, it works from day one with few labels and can read unstructured text like descriptions, names, addresses and item titles, which tree-based models cannot.

Classical models handle the opposite case: logistic regression and XGBoost-compatible gradient-boosted trees learn a company's own outcomes precisely from its own numeric data, at no per-decision model cost. The two can be stacked, with Grayson's answers becoming features alongside numeric fields in a traditional model, while experiments show which combination actually wins on that dataset.

Getting it running

Kairos is available as a hosted cloud service at kairos.finic.ai, or self-hosted. Developer surfaces include a TypeScript SDK, a Python SDK, a CLI and a plain HTTP API. Self-hosting requires Node.js 22, pnpm 10 and a Grayson API key from Finic's portal; the server starts its own Postgres instance, and teams can deploy with docker compose. Internally, the server is Hono on Node.js with Postgres accessed through Drizzle, and the console is React and Vite.

Integrations listed in the repository cover Shopify — deciding each order, holding risky ones and learning from chargebacks — plus inbound webhooks for providers such as Stripe.

Why it matters

Fraud platforms from Sardine, Sift and SEON are proprietary and typically tuned by vendor teams or in-house ML engineers, and most ML fraud models need thousands of labeled chargebacks before they are useful. Kairos bets on two shifts at once: that coding agents can now handle the continuous tuning loop on a company's own data, and that a decision model can cover the cold-start period before those labels exist.

The open-source angle also carries weight in a sensitive domain: teams can audit exactly which signals produced each decision and self-host rather than shipping event data to a third party. Caveats apply. Grayson remains a proprietary component behind an API key, so the open-source license covers the decisioning platform rather than the full model stack, and the launch claims come from the project's own materials rather than independent evaluation.