Ojuri
What it is
Ojuri is a free, MIT-licensed, self-hosted fraud detection engine for fintechs, payment gateways, and mobile-money platforms. It evaluates transactions and returns accept, review, or decline decisions. Applications can call the POST /v1/predict API, while the Node SDK provides typed requests, retries, and webhook signature verification. Real-time scoring uses an XGBoost model served through ONNX Runtime on the authorization path. A pattern-analysis agent follows account, device, and payee relationships over time to identify fraud rings and mule networks. Decisions include reason codes and the model data used, with records kept in an audit log. The Sentinel dashboard offers a review queue, decision stream, model registry, and analyst overrides that can become training data. Model-learning tools monitor drift, retrain on verified labels, and check evaluations before promotion. A separate investigation agent can create reports for declined transactions using an optional self-hosted language model. Services run in the operator’s infrastructure, keeping transaction data within that boundary.
For and against
For
- Free MIT-licensed plan
- No per-call fees or rate limits
- Real-time XGBoost scoring
- Fraud-ring and mule-network analysis
- Self-hosted data boundary
Against
- Docker and Node 20 required for quickstart
- Optional investigation model downloads about 7.6 GB
- Self-hosting requires operator infrastructure
- Not a compliance certification
The verdict
Ojuri provides transaction decisions, relationship analysis, investigation tools, and model management in a self-hosted package. Its free license and data-boundary design fit teams that can operate the required infrastructure.
Who it suits
Fintechs, payment gateways, and mobile-money platforms that need self-hosted transaction scoring. It suits teams requiring auditability and control over transaction data.
Plans and prices
Free
Self-hosted deployment · no rate limits · customer-operated infrastructure
Read 15 Oct 2026 · ojuri.io
Ojuri
MIT licensed · no per-call fees · no rate limits
Read 8 Oct 2026 · ojuri.io
On the record
- Analyst dashboard
- Sentinel includes a review queue, decision stream, model registry, and analyst overrides that can become training data.ojuri.io · read 8 Oct 2026
- Analyst workflow
- Sentinel provides a review queue, live decision stream, rule editor, model registry, audit log, investigations, and user administration.ojuri.io · read 15 Oct 2026
- Architecture
- The system uses four agents: real-time detection, pattern analysis, model learning, and fraud investigation.ojuri.io · read 15 Oct 2026
- Audit trail
- Ojuri records each decision and the feature snapshot used at inference before returning the response.ojuri.io · read 15 Oct 2026
- Benchmark result
- The repository reports 34% fraud detection before retraining and 98.8% after one label-driven retraining round, with 1.1% false positives in its simulation.github.com · read 15 Oct 2026
- Compliance
- Ojuri says self-hosting can help with data residency requirements including GDPR, NDPR, and CBN, and explicitly says it is not a compliance certification.ojuri.io · read 8 Oct 2026
- Data handling
- Ojuri says its services run in the operator’s infrastructure and transaction data does not leave that boundary for scoring, training, or investigation.ojuri.io · read 8 Oct 2026
- Data residency
- The maker states that scoring, training, and investigations run inside the customer’s infrastructure without transaction-data egress.ojuri.io · read 15 Oct 2026
- Deployment
- The quickstart uses Docker Compose and requires Docker and Node 20; the investigation model is optional and downloads about 7.6 GB of weights when enabled.ojuri.io · read 8 Oct 2026
- Explainability
- Responses include reason codes, feature contributions, observed values, model version, threshold, and audit ID.ojuri.io · read 15 Oct 2026
- Extensibility
- Custom pre-model and post-model rules can be edited without restarting and normally update within 30 seconds.github.com · read 15 Oct 2026
- Feature extensibility
- The default model exposes 64 features and supports defining additional signals without code changes.github.com · read 15 Oct 2026
- FIA resource limit
- The optional investigation service downloads about 7.6 GB of model weights and needs about 16 GB of free RAM.github.com · read 15 Oct 2026
- Fraud-ring detection
- Pattern analysis identifies shared-device clusters, relationship groups, and emerging transaction rings.ojuri.io · read 15 Oct 2026
- Fraud rings
- The Pattern Analysis Agent tracks account, device, and payee relationships over time to identify fraud rings and mule networks.ojuri.io · read 8 Oct 2026
- Integrations
- The stack uses Kafka, Redis, PostgreSQL, ONNX Runtime, Docker Compose, NGINX, and Prometheus/Grafana components.github.com · read 15 Oct 2026
- Investigations
- The Fraud Investigation Agent creates reports for declined transactions using a self-hosted language model and runs on a separate path from live scoring.ojuri.io · read 8 Oct 2026
- Latest release
- The homepage changelog identifies v1.7.0, dated September 28, 2026, as the current release shown there.ojuri.io · read 15 Oct 2026
