AI Fraud Detection for a Neobank Startup
Graph + behavioral fraud scoring that cut false declines while catching account-takeover rings for an early-stage neobank.
Key Details
| Challenge | Rules-only fraud checks blocked good users and still missed coordinated account-takeover attempts. |
|---|---|
| Solution | A real-time fraud scorer combining device, velocity and graph features with an investigator console. |
| Technologies | Python, Kafka, Redis, PostgreSQL, AWS |
Technologies used
Client background
A neobank startup shipping cards and P2P transfers hit a wall: static rules either annoyed legitimate users with declines or let mule rings through. Fraud ops lived in Slack threads with no shared case model.
Key challenges
- False declines hurt activation and support costs for a brand still earning trust.
- Coordinated rings shared devices and funding patterns rules could not see.
- No real-time feature path — batch jobs were too slow for card authorizations.
- Investigators lacked a prioritized queue with evidence packs.
Product screenshots
What we built
- Streaming features for device fingerprint, velocity and peer-graph risk.
- Online fraud score service with kill-switches and reason codes.
- Investigator console with case notes, linked entities and disposition outcomes.
- Feedback loop so confirmed fraud retrains weekly challenger models.
Project team: 6 engineers across AI/ML, backend and product engineering — delivery over 14 weeks for a startup team that needed to ship, not slide decks.
Architecture design
Neobank
Customers
Cards & P2P transfers
Card authorisations
Real-time
Devices & sessions
Fingerprints
Investigators
Fraud ops
Applications
Investigator console
Prioritised queue
Case notes
Linked entities
Dispositions
Outcomes
AI & ML
Online fraud score
Reason codes
Kill-switches
Safe rollback
Peer-graph risk
Linked rings
Challenger retraining
Weekly
Data & streaming
Streaming features
Kafka
Velocity & device signals
Real-time
Feature cache
Redis
Case & feedback store
PostgreSQL
Cloud & data services
Streaming
Kafka
Feature streams
Redis
Feature cache
Data
PostgreSQL
Cases & feedback
Cloud
AWS
Hosting
Docker
Containers
Technology stack
Backend & APIs
Python
Data
KafkaRedisPostgreSQL
Cloud & DevOps
AWSDocker
How we delivered
01
Baseline
Measured false decline vs. fraud loss and mapped the hottest attack patterns.
02
Features
Stood up streaming feature jobs with latency budgets for auth time.
03
Score
Shipped v1 scorer with human review for mid-risk band.
04
Operate
Trained fraud ops and locked SLAs for model and rule changes.
Business impact
-35%False declines
+2.1×Fraud catch rate
<80msScore latency