Products / In-process AI

jitNN

Train and serve AI models inside your Java platform, as fast as PyTorch or faster.

Talk to us about jitNN jitNN
At a glance● data stays put
2.4×faster ResNet-50 inference than PyTorch (A100, batch 1)
Sameresults as PyTorch, checked operator by operator and model by model
0code changes for existing DJL applications
1runtime for data, training and serving
The problem

AI models that live outside your Java platform

Your transactions, customers and events live in Java systems, but models are trained and served in Python. Every model needs an export pipeline, a second runtime to run and secure, and a hand-off between data-science and platform teams. Older Java ML libraries avoid that, but cannot use a modern GPU.

What Beestream does

Models where your data and services already are.

jitNN trains and serves deep-learning and classical machine-learning models inside the JVM, on the GPU. Bring the ONNX, TensorFlow and Hugging Face models you already have, or keep your DJL code as it is, and run them in the same process as the services that use them.

What you get
  • Bring your models: ONNX, TensorFlow and Hugging Face GGUF files load directly
  • Keep your code: DJL applications run unchanged
  • Deep learning and classical ML in one library, from ResNet, BERT and GPT-2 to gradient boosting and random forests
  • Results you can trust: checked against PyTorch, scikit-learn and ONNX Runtime
  • Grows with you: training across several NVIDIA GPUs, and a CPU fallback where there is no GPU
ResNet-50 inference, batch 1, FP32 · milliseconds, lower is better
PyTorch2.14, CUDA
8.12 ms
ONNX Runtime1.30, CUDA
4.05 ms
jitNNwith vendor libraries
3.40 ms
NVIDIA A100 80GB · host input to host output · median of 50 calls
Where it pays off

Built for the decisions your business runs on.

Fraud and risk scoring

Score every transaction in the Java service that processes it, with no call out to a model server.

Recommendations

Train and refresh recommendation models on the event data your platform already holds.

Documents and images

Classify, tag and search documents and images with BERT- and ResNet-class models.

Retiring legacy Java ML

Move DL4J- and DJL-era code onto GPUs without a rewrite in Python.

The Beestream platform

Start with one product. Add the rest on the same GPUs.

jitNN is one part of the Beestream platform. Once it is in, the other products run on the same GPUs, under the same support, with no new stack to adopt.

7.3×Apache Iceberg GPU accelerationfaster Iceberg compaction UDFsSpark RAPIDS GPU accelerationcompiled to GPU kernels 19×JVector GPU accelerationfaster vector index builds 1.4×jitLLMllama.cpp prompt speed 13×jitCastfaster forecasts
Contact

Tell us what is costing you CPU time. We'll show you the GPU path.

LinkedIn/company/beestream-ai X@beestream_ai Bluesky@beestream.ai GitHubgithub.com/beestream-ai