Ojuri vs. traditional SaaS fraud detection
Scaling a fintech, payment gateway, or mobile-money platform usually means choosing between an expensive black-box SaaS vendor and months building a rules engine from scratch. Ojuri is a third option: an open-source, self-hosted, multi-agent fraud engine that runs inside your own infrastructure — no per-call fees, no data egress.
| SaaS fraud API | Ojuri | |
|---|---|---|
| Deployment | Third-party SaaS API | Self-hosted, single-command Docker Compose |
| Per-transaction latency | External network round-trip per call | In-process · ≈6ms server-side (reference benchmark) |
| Data residency | Payloads leave your network | Stays inside your boundary |
| Pricing | Per-call API fees | Free — MIT licensed |
| Model updates | Vendor-controlled, often manual | Drift-aware automated retraining |
| Explainability | Often a black box | Reason codes + LLM investigation reports |
Scoring runs in your process, not across a network
SaaSA hosted fraud API adds a network round-trip to every authorization, so scoring latency depends on a third party’s data center and your link to it.
OjuriOjuri serves a compiled XGBoost model via ONNX Runtime inside the detection service — scoring is in-process, with no third-party hop. The project’s reference benchmark measures p99 ≈ 6ms server-side on a single developer workstation; these are reference values, not SLA targets, so re-measure on your own hardware.
Payloads never leave your infrastructure
SaaSSending PII and transaction payloads to a third-party cloud means customer data leaves your network — a data-residency concern under regimes like NDPR, CBN, and GDPR.
OjuriOjuri runs entirely inside your network boundary under an MIT license. Customer profiles, payloads, and risk metrics stay on your own servers.
Models retrain on real labels, not static rules
SaaSLegacy rules drift out of date and drive false positives; updating models often means slow, manual pipeline work.
OjuriThe Model Learning Agent watches F1 and PSI, retrains on reviewer ground-truth labels, and gates promotion behind statistical checks before anything reaches production.
Four specialized agents, decoupled by Kafka
Unlike a monolithic scoring API, Ojuri splits fraud detection across independent agents so a failure in one never blocks a payment in another.
- rdaReal-Time Detection AgentReal-time fraud detection on the authorization hot path
- paaPattern Analysis AgentVelocity checks and fraud-ring detection, off the hot path
- mlaModel Learning AgentDrift-aware retraining with shadow-to-active promotion gates
- fiaFraud Investigation AgentExplainable LLM fraud investigations, on a separate path