Rimay Quantum Feature Extraction
Quantum features for your tabular data. Same models, new headroom.
Published in Mayo Clinic Proceedings · Scientific Reports · Physical Review Research
Built for hard data
Scarce
Noisy
Imbalanced
Drift
Quantum features, into your Python pipeline
Upload a tabular dataset to a Hub data pool, run Rimay, and it writes back quantum features as NumPy arrays for scikit-learn, XGBoost or PyTorch. The same workflow runs unchanged on the free IBM Aer simulator and on IBM Heron hardware.
import json, numpy as np
from qhub.api.platform.client import HubPlatformClient
from qhub.service.client import HubServiceClient
# 1. Upload your tabular data to a Hub data pool (the file must be data.json).
pools = HubPlatformClient(api_key=QHUB_ACCESS_TOKEN, organization_id=ORG_ID).data_pools
dataset = {
"training_tabular_data": X_train.to_dict(), # <= 15 features, >= 20 rows
"training_target_data": y_train.to_dict(),
"test_tabular_data": X_test.to_dict(),
"test_target_data": y_test.to_dict(),
}
src = pools.create_data_pool(name="Rimay input")
pools.add_data_pool_file(id=src.id, file=("data.json", json.dumps(dataset).encode()))
dst = pools.create_data_pool(name="Rimay output")
# 2. Extract quantum features on the free IBM Aer simulator.
client = HubServiceClient(
access_key_id=ACCESS_KEY_ID, secret_access_key=SECRET_KEY,
service_endpoint=SERVICE_ENDPOINT,
)
client.run(request={
"input_data_pool": {"id": src.id, "ref": "DATAPOOL"},
"output_data_pool": {"id": dst.id, "ref": "DATAPOOL"},
"num_shots": 500,
"num_runs": 1,
}).result()
# 3. Single-body expectations + two-body correlations, ready for any classifier.
Xq_train = np.load("1_Xq_train_0.npy")
model = XGBClassifier().fit(Xq_train, y_train)QPU during training only, classical deployment
Off-line surrogate framework · arXiv 2605.19801
Quantum extracts
A quantum processor encodes a representative subsample of your data into many-body spin dynamics and reads out single-body expectations and two-body correlations.
Surrogate learns
A lightweight classical model learns the quantum-induced representation. Quantum feature mappings are stable across backends, so it generalises reliably.
Deploy classical
Every prediction runs on classical hardware: microsecond latency, no quantum queue, and the same MLOps and procurement as any classical model.
What leaders are saying
“Kipu Quantum's quantum feature extraction is a great example of a cost-effective way to run hybrid QML workflows […] with our quantum hardware delivering accurate results across a wide range of applications.”
Scott Crowder
Vice President, IBM Quantum Adoption
“Kipu's off-line surrogate framework achieves economic quantum advantage, capturing the 2–3% accuracy gains of a quantum processor while running inference entirely on classical hardware […] actively applied to high-volume enterprise problems.”
André König
CEO, Global Quantum Intelligence
“Kipu's quantum feature surrogate framework marries quantum-derived representations with the classical infrastructure enterprises already trust: measurable accuracy gains, zero quantum dependency at inference. We are ready.”
Rika Nakazawa
Chief Commercial Innovation, NTT DATA
“Through the Kipu Quantum Hub we are optimizing classical models in image classification for predictive maintenance […] using thermographic drone imagery for early detection of issues in our energy parks.”
Estela Vilches
Head of Digital Innovation, MOEVE
“The scope of this technology is intentionally broad and industry-agnostic […] letting enterprises leverage the computational advantages of quantum systems across their entire portfolio of data-intensive challenges today.”
Aaron Kemp
Senior Director Quantum Research & Enterprise Innovation, KPMG US
The research behind it
Peer-reviewed and preprint work on quantum feature extraction and quantum-enhanced machine learning from Kipu Quantum.
Questions, answered
What is Rimay?
Rimay is Kipu Quantum's quantum feature extraction service. It transforms classical tabular data into richer quantum feature representations that boost the accuracy of standard machine learning models. Rimay runs as a managed service on the Kipu Quantum Hub, where it executes on IBM quantum hardware, and as a free simulator service for development and testing.
What does quantum feature extraction actually produce?
For each row of your dataset, Rimay returns two kinds of quantum-derived features: single-body expectation values, one quantum-transformed value per feature, and two-body correlations between pairs of features whose qubits are physically coupled in the backend's topology. These are concatenated into a new feature matrix you append to or swap in for your classical features, then train any classifier on top.
How does Rimay extract features?
Rimay encodes your classical features into the dynamics of many-body spin Hamiltonians on a quantum processor, then reads out expectation values of low- and higher-order observables. This captures statistical dependencies and higher-order correlations that classical preprocessing struggles to reach. Feature-to-qubit mapping is hardware-aware: each feature gets one qubit, and the feature pairs read out as two-body correlations are the ones whose qubits the processor can couple directly.
Do I need a quantum computer in production?
No. With Kipu Quantum's off-line quantum feature surrogate framework, the quantum processor runs only during a targeted training stage on a small representative subsample, as little as 20 percent of the data. A lightweight classical surrogate then learns the quantum-induced representation and runs every prediction. Deployment is fully classical: microsecond inference latency, no quantum queue, and the same MLOps and procurement profile as any classical model. This delivers at least five times fewer quantum executions for the same accuracy.
What hardware and limits does Rimay support?
The free Rimay Feature Extraction Simulator runs on the IBM Aer simulator with up to 15 features and 1,000 samples, ideal for prototyping. The managed Rimay service runs on the IBM Aer simulator up to 15 features or on IBM Heron quantum processors (ibm_kingston, ibm_torino, ibm_fez, ibm_marrakesh) up to 133 or 156 qubits, with up to 20,000 samples. You control the number of shots and runs per execution.
How does Rimay fit into my existing ML pipeline?
You upload your training and test data to a Hub data pool, run Rimay with a chosen backend, and the service writes the quantum features back as NumPy arrays. From there the features drop straight into your existing scikit-learn, XGBoost, or PyTorch workflow. An optional inference mode applies the same quantum transformation to new, unlabeled data.
How is Rimay different from Iskay and Miray?
Rimay, Iskay, and Miray are distinct Kipu Quantum services. Iskay and Miray are the same optimization engine, a QUBO and HUBO solver powered by bias-field digitized counterdiabatic optimization. Rimay is a different capability: it does not solve an optimization problem, it generates quantum features from your data to improve machine learning accuracy. The three are designed to be combined on the Kipu Quantum Hub.