Fraud detection with machine learning
Short answer
SwankyForge builds fraud and abuse detection that scores transactions, orders or accounts in real time, explains each decision, and runs alongside your existing rules until it is proven. The goal is fewer false alarms at the same catch rate, measured on your data. Projects are scoped after a short discovery, within a typical range of €7k–€42k.
- Order 88412 · returning customer0.08
- Order 88413 · new device, usual address0.21
- Order 88414 · new device, new address, high value0.87
- Order 88415 · returning customer0.05
Flagged: new device + new address + order value 6× the customer's usual
Sounds familiar?
- Rules catch only the fraud they already know.
- Analysts spend the day on false alarms.
- Good customers get declined.
- New attack patterns go unnoticed for weeks.
- Nobody can explain why an order was blocked.
What we build
Real-time scoring
Every transaction, order or sign-up scored inside your existing flow.
Costed thresholds
Cut-offs set by what each kind of mistake costs you, not by a default.
Reasons for every decision
The factors behind each score, readable by analysts and support.
Drift alerts
A warning when behaviour shifts, before losses show up.
Analyst queue
Cases ordered by risk and value, with the evidence attached.
Gradual rollout
Runs next to your rules first; takes over only where it is better.
How a project runs
- 1
Count the cost of mistakes
What a missed fraud costs, and what a wrongly declined customer costs. Both set the target.
- 2
Build the history
Past decisions, chargebacks and reviews become labelled examples.
- 3
Score and explain
A model scores each event in real time and lists the reasons.
- 4
Shadow your rules
It runs silently next to the current rules until it catches the same fraud with fewer false alarms.
- 5
Monitor and retrain
Drift alerts, weekly reports and a rollback you control.
What you get
- Scoring service inside your payment or order flow
- Threshold policy based on the cost of each error
- Analyst queue with reasons per case
- Drift and performance monitoring
- Code, models and runbooks in your repository
Typical price
€7k–€42k
Typical timeline
Scoped on discovery
From the published rate card. The final price is fixed once the scope is agreed.
Estimate your projectQuestions
Will machine learning replace our fraud rules?
Not at first. The model runs alongside the rules and takes over only where it catches the same fraud with fewer false alarms. Some rules stay because they encode policy.
How fast is real-time scoring?
Fast enough to sit inside checkout or payment authorisation. The response-time budget is agreed up front and monitored.
Can you explain why a transaction was declined?
Yes. Each score comes with the factors behind it, so analysts and support can explain a decision to a customer.
How much labelled fraud do we need?
Confirmed fraud is always scarce. We combine chargebacks, past reviews and anomaly signals, and measure what is achievable on your history first.
Why is there no fixed price?
Integration points and data quality vary most between companies. A short discovery sets the scope and the price.
Tell us about the task.
The first call is free. If this system is not the right answer for your numbers, we say so.