atrosa

How it works

The run

0102030406101323303643515964727584904 exports100 iterations, each scored1 rule kept, hashed
  1. You send an export

    CSV or JSON, in as many files as you have. Columns and timestamps are detected on arrival. The agent works on the files, so it needs no connection to your production systems.

  2. The agent writes, runs and scores

    It writes a detection in Python, runs it in a sandbox on your data, scores the result without labelled cases and rewrites it. One run is up to 100 of these iterations.

  3. You keep what scored

    Your team reviews the rules that scored and chooses which ones run on new data. Each alert reaches Slack or a webhook with a draft report attached.

The agent runs on one GPU with a local model, or on a cloud model you choose, and your data is used for your rules alone. A model trained on your data, kept in your environment, is in development. See Fine-tuning.

Results

One run on one operator's data.

A cross-border remittance operator sent four exports, about 2.5 million rows, with no labelled cases. The agent ran 100 iterations in 5 hours 6 minutes, and the operator checked the five flagged accounts by hand.

accounts read
301,099
flagged
5
confirmed fraudulent
4
known case missed
1

What the kept rule keyed on

Reversal rate27 times the peer median
Transactions per account-day3.08, against a population median of 0.04
Sourcesflagged accounts present in all four
Score per iteration, one runbest 53 of 100 at iteration 84
1flat marks: crashed or rejected100

Across 51 runs on this data the agent wrote 2,037 rule versions, of which 12 were kept. There is no public benchmark yet.

Cost

What it costs, and what the same volume costs today.

Open sourceFreeThe public detections, helpers and tests, under the MIT licence.
Self-servefrom $5,000 a yearHosted by us. You send exports and the agent runs on them.
Platformfrom $50,000 a yearRuns inside your environment, with draft reports, and fine-tuning when it ships.
First run on your exportNo chargeOne full run on one export. You keep every rule it produces.

Transaction-monitoring platforms

Quoted to us for the same volume, 1.2 million transactions a year, at $42,000 to $107,573 a year on one- to three-year terms billed up front, after weeks of procurement and setup.

By hand

An analyst or consultant writes rules in a vendor’s builder, on the part of your data the builder can see, and revises them on a quarterly cycle.

Who it's for

Risk and compliance teams of two to five
A fintech, remittance company or neobank whose data sits in Stripe, Postgres, the ledger and the logs, and whose rules live in a vendor's builder or a spreadsheet.
Hosted
Institutions that keep detection logic in-house
The rules, and the model that writes them, run on your hardware, and your engineers review each rule as code before it goes live.
Your environment
Not a fit
Real-time blocking at checkout, and teams that cannot produce a data export. ATROSA runs alongside your case management and sanctions screening, which stay as they are.

Questions

Do we need labelled cases?
Rules are written without them. If you have confirmed cases, they raise the score of the rules that catch them.
Does it touch production?
It reads exports only. Alerts, and suspend calls if you enable them, go out through a webhook you configure.
Does the agent run unsupervised?
It writes and tests rules on its own. A rule runs on new data only after your team adds it to the active set.
Is a model trained on our data?
Today a general model writes the rules from your exports. A model trained on your data alone, kept in your environment, is in development.
Who files the suspicious activity report?
Your compliance officer does. ATROSA drafts the report from the rule and the alert.
How long does a run take?
The reference run took 5 hours 6 minutes for 100 iterations. A first run returns results within 24 hours of our receiving the export.
How is our data handled?
On a local deployment everything stays on your machines. Rules are hashed, so an edited file will not load until you approve it again. ATROSA is not yet SOC 2 certified, and we send the current security documentation before you send anything.