FinTech

Real-time fraud detection inside the transaction flow

  1. Problem

    A payments platform needed to score every transaction for fraud in real time – but routing traffic through a separate Python scoring service added latency, infrastructure, and a second stack to operate.

  2. Solution

    Deep learning models built with Deep Netts run directly inside the Java transaction pipeline. Scoring happens in-process, with no network hop and no additional runtime to deploy or monitor.

  3. Outcome

    Fraud decisions now land in milliseconds, inside the same JVM that processes the payment – and the fraud team iterates on models with the tools the platform team already runs.

payment-service · fraud scoring (in-process)
0
Scored
0
Approved
0
Blocked
0
Last score
Step 1 / 5
Payments

Every payment passes through the platform's Java service. Fraud is a small fraction of them, and its patterns keep changing – rules alone are easy to bypass.

30%
lower fraud losses (the case)
ms
decisions, in the same JVM
0
extra Python infrastructure

Case: payments platform, not named – 30%, ms and 0 from the case. Features, model, 0.70 threshold, benchmark and test results: Deep Netts talk “AI in Java for the Financial Industry”, Amsterdam JUG at ING, Oct 2026 (an example model, not the customer's). Payments drawn are illustrative.

Enterprise IT

Forecasting cloud resources before the bill arrives

  1. Problem

    An enterprise platform team was over-provisioning cloud capacity to stay safe, because usage forecasts were coarse and reactive – and the cost of that safety margin kept growing.

  2. Solution

    Deep Netts models forecast resource demand from historical usage, running natively inside the team's JVM-based tooling. Predictions feed autoscaling and capacity planning directly, with no separate ML pipeline.

  3. Outcome

    Provisioning now follows predicted demand instead of worst-case guesses, cutting spend while improving stability during load spikes.

capacity-planner · demand forecast
–
Month
–
Demand vCPU
–
Capacity vCPU
–
Idle
Step 1 / 5
Worst case

Without a usable forecast, capacity is sized for the worst case and kept there all year. Most of it is paid for and sits idle.

15–25%
cloud cost reduction (the case)
JVM-native
forecasts inside the existing tooling
No ML pipeline
predictions feed autoscaling directly

Enterprise IT customer, not named. The 15–25% figure is from the case; demand, forecast and capacity drawn here are illustrative.

Scientific Computing – Jefferson Lab

Predicting particle trajectories with neural networks

Read the full story
  1. Problem

    Every event in the CLAS12 detector leaves many segments in the drift chambers – from real particles and from background. The conventional algorithm has to try their combinations, tracking takes 80–90% of event processing time, and at high beam intensity good tracks get lost among the candidates.

  2. Solution

    Two neural networks built with the Deep Netts community edition, in Java, inside the lab's Java reconstruction software: a multi-layer perceptron that labels each track candidate a positive, negative or false track, and an auto-encoder that fills in a segment a chamber missed.

  3. Outcome

    On real data, 15.6% more reconstructed tracks – over twice as many with a missing segment – and 35% faster tracking. Physics analyses gain 15–35% more events (G. Gavalian et al., arXiv:2202.06869).

CLAS12 · drift-chamber track reconstruction
0
Segments
0
Candidates
0
Rejected
0
Tracks found
Step 1 / 6
Collision

A 10.6 GeV electron beam hits a 5 cm liquid-hydrogen target. This event has an electron and two pions – one of the reactions the paper analysed.

+10–15%
tracking efficiency
35%
faster tracking
+15–35%
events in two- and three-particle reactions

Source: G. Gavalian et al., CLAS12 Track Reconstruction with Artificial Intelligence (arXiv:2202.06869) – detector layout, network, two-pass workflow, Table 1, run conditions. The event drawn is illustrative.

Research – University of Minnesota

Deep learning for drug discovery, in Java

Read the full story
  1. Problem

    High-throughput screening produces tens of thousands of data points in one session. Finding the compounds that cause the desired change in a protein calls for deep learning – but tools that depend on native libraries are hard to run on specific hardware, and QSAR methods need structural data such as SMILES that many library compounds lack.

  2. Solution

    Oscar Bastidas, Ph.D., analyses the screening data with Deep Netts feed-forward networks in Java, trained on the physical and chemical properties already in the compound library, with hyperparameters tuned to the data. His study of spike-protein dynamics with Zoran Ševarac, analysed in Java, was published in Scientific Reports (2024).

  3. Outcome

    Hit identification runs entirely on the JVM – no native dependencies, no QSAR input, and in the language the researcher already works in.

protein dynamics · compound screening
0
Frames
0
Top frequency
0
Screened
0
Hits
Step 1 / 5
Protein dynamics
Simulation

The study simulated the SARS-CoV-2 spike protein (PDB 6VSB) in water for 200 ns, one frame per nanosecond. From every frame the backbone angles φ and ψ of each residue are read – here φ of one residue.

200 ns
simulated, one frame per ns (PDB 6VSB)
23–63 MHz
favoured φ/ψ oscillations, top 42.97 MHz
Pure JVM
deep learning with no native libraries

Sources: O. H. Bastidas & Z. Sevarac, Time dependent dihedral angle oscillations of the spike protein of SARS-CoV-2, Scientific Reports (2024); O. Bastidas, Ph.D., Deep Learning in Drug Discovery (2021). The protein fragment, compounds and values drawn are illustrative.

Have a workload like these?

Tell us about your system – we'll show you how Deep Netts fits inside it.