Our work
The thinking behind Node16.ai, what we have built and what our experiments show.
Public white paper · v0.1 · 2 pages
One introduction to the business case and architecture: value for a first bank, Sonar and Investigator, the data boundary, oversight and the route to Collaborator.
Download the white paper ↓Presentation · NVIDIA FLARE Day 2026
Public slides updated 23 September 2026
Our work on federated graph learning was presented at NVIDIA FLARE Day 2026. The talk covers learning across banks that keep their data separate, pseudonymous counterparties and lessons from building on sovereign GPU infrastructure.
The slides include our verified synthetic results. The results section below sets out the comparison and the investigation-workload trade-off.
Published research · HICSS-59, 2026
Research on designing agentic compliance workflows with bounded roles, traceable decisions and explainability. The paper draws on a fintech prototype and examines how accountable governance can support automation across compliance work.
Read the published paper ↗Proceedings of the 59th Hawaii International Conference on System Sciences.
Research article · Under peer review
A research paper on the governance of distributed AI systems is currently under peer review.
Further details will be shared after the review process.
Technical overview
Detailed architecture, methods and evaluation evidence.
Our experiments · Results verified 17 September 2026
A useful detector has to balance what it finds with what a team can investigate. In our synthetic experiments, shared learning improved detection for all five simulated banks. These results come from synthetic data with a dense transaction network between the banks; real-bank performance is what we test next with partner banks.
At 1,000 alerts per million transactions
98.4–99.5%That is approximately 5–16 false alerts per 1,000 alerts across the five simulated banks.
This queue recovered 23.2–35.5% of all labelled suspicious transactions. Selecting a smaller queue leaves more suspicious activity outside it.
| Transactions flagged | Precision | Recall |
|---|---|---|
| 0.1% | 98.4–99.5% | 23.2–35.5% |
| 0.5% | 42.7–64.4% | 75.2–77.4% |
Precision is the share of alerts labelled suspicious. Recall is the share of all labelled suspicious transactions recovered. These are transaction alerts, before grouping them into cases or reports. What this means for your team →
Sonar provides the detection signal. Investigator brings it together with graph exploration, KYC context and case evidence, so the investigator can assess the activity and prepare the next step. Investigation and reporting decisions stay with the bank.
The run used four retained investigation windows per simulated bank, drawn from our synthetic corpus of approximately 4 billion transactions and 4 million customers, covering 187 money-laundering typologies. The corpus scale is distinct from the selected training and evaluation windows. All five simulated banks ran on one sovereign compute host.
Local and shared graph models used matched bank-local inputs. The shared model's precision–recall ranking improved over the local model for all five simulated banks. Thresholds were selected on development validation; some input signals reflect patterns built into the synthetic generator. The public presentation slides 5 and 10 report the same comparison, and the technical overview will carry the training rounds, hardware, metrics and thresholds.
Further reading
Selected external work on collaborative detection, synthetic data and accountable AI. These publications report their own methods and findings.
A research architecture combining horizontal and vertical federated learning for financial-crime detection.
A synthetic five-site study by NVIDIA and financial-sector researchers, exploring shared learning and sample-level differential privacy.
Self-supervised graph learning for AML, connecting transaction patterns with analyst review.
A financial-fraud study using bank data in a controlled research environment. Its operating setup differs from bank-local deployment.
BIS Innovation Hub research into collaborative analytics, machine learning and privacy-enhancing technologies for AML.
FCA research on generating synthetic data to develop and test money-laundering detection approaches.
Practical financial-sector guidance on AI governance, oversight and lifecycle risk management.