NEXT GEN
CREDIT SCORING

A production-grade Hybrid Quantum Machine Learning pipeline running on the IQM Garnet Processor.

Key Achievements

99%
Cost/Latency Reduction
Via Batch Execution (Track A)
71.5%
Hardware Inference Accuracy
Sim-to-Real Transfer (Track B)
89%
Loss Reduction on QPU
Verified Learning (Track A)
80%
Full Simulation Accuracy
Theoretical Limit (QSVC)

"Two Tracks, One Goal" Benchmark

We stress-tested our approach via two parallel tracks to ensure robustness: Track A optimized for real hardware constraints, while Track B validated the theoretical upper bounds in simulation.

TRACK A: HARDWARE OPTIMIZATION

PRODUCTION (LOQI)

Focus: Speed & Real Noise Resilience

Backend IQM Garnet (20q QPU)
Architecture 5-Qubit VQC (High Fidelity)

KEY ACHIEVEMENT

99%
Cost/Latency Reduction
89% Loss Reduction on Hardware
  • ✅ Batch Execution Middleware implemented
  • ✅ Topology-Aware Transpilation
  • ✅ Verified Hardware Convergence

TRACK B: SCIENTIFIC VALIDATION

VALIDATION (QRISP)

Focus: Theory & Sim-to-Real Transfer

Backend IQM Sirius (20q QPU)
Architecture 4-Qubit VQC (Compact)

KEY ACHIEVEMENT

71.5%
Hardware Inference Accuracy
Validated on 900 Real Samples
  • ✅ Weighted MSE for Imbalance
  • ✅ Pure Simulation Baseline
  • ✅ Validated Model Stability

Interactive Visualizations

Technical Innovations

Batch Execution Optimization

Reduced API overhead by 99% through circuit batching. Submitted 100 circuits per job instead of individual executions, enabling full training within credit constraints.

Topology-Aware Transpilation

Dynamically extracted IQM Garnet's coupling map and applied optimization level 3 to insert SWAP gates automatically, ensuring circuit validity on real hardware.

Physics-Informed Preprocessing

Implemented 0-to-π Scaling strategies to maximize Angle Embedding expressivity on the Bloch sphere, preventing phase redundancy and "feature crowding".

Class Imbalance Mitigation

Integrated a Weighted MSE cost function (Track B) to penalize minority class errors, ensuring the model doesn't just predict "safe" non-defaults.

Unweighted Cost Matrix
Weighted Cost Matrix

Left: Unweighted (High Accuracy, Low Recall). Right: Weighted (Balanced, High Recall).

Quantum Circuit Architectures

Variational Quantum Classifier (VQC)

Hardware-efficient ansatz with circular entanglement pattern. Angle embedding via RX gates.

VQC Circuit

Quantum Neural Network (QNN)

Data re-uploading architecture. Interleaves feature encoding with trainable rotations across multiple layers.

QNN Circuit

Quantum Support Vector Classifier (QSVC)

Fidelity quantum kernel using inversion test. Computes kernel matrix via quantum interference.

QSVC Circuit

Team LOQI

Built for the Banco Santander Quantum Challenge 2025. Our mission: Demonstrate production-ready quantum machine learning on real hardware with rigorous engineering standards.

Team LOQI
David Blanco
Fathin Dosunmu
Iker Rodriguez
Alejandro de los Santos Bravo
Hüseyin Umut Işık
Qrisp
IQM Garnet 20q
Plotly
Scikit-Learn
XGBoost
Python
View Repository on GitHub