Renee Julian S.L.

Quantum Machine Learning Consultancy Barcelona

We build quantum machine learning systems that survive contact with real hardware.

Renee Julian S.L. designs, implements and benchmarks quantum learning pipelines — and reports the results, including the ones that don't win.

See the work quantum@reneejulian.com
DisciplineQuantum ML · NISQ
StackQiskit · PennyLane · IBM Heron
BasedBarcelona, ES
01

Selected work

Three machine learning projects we built, measured and published — with the measured numbers attached.

Active · 2026 PennyLane · LightGBM · IBM Quantum

Hybrid quantum–classical ensemble for financial prediction

An 8–12 qubit data-re-uploading QNN, angle-encoded from a PCA-reduced feature basis, ensembled with a LightGBM baseline and scored on the NumerAI tournament. Trained with SPSA rather than backpropagation, calibrated against the live noise profile of IBM's ibm_marrakesh Heron r2 processor.

The honest result: the quantum arm is 86–99.7% orthogonal to the classical baseline — genuinely independent signal — but not yet strong enough to lift ensemble correlation above LightGBM alone. It remains in live tournament submission while the feature map is reworked.

Source
Validation
649 eras, walk-forward
LGBM baseline
ρ = 0.0151
QNN standalone
ρ = −0.003
Orthogonality
86 – 99.7%
Tournament MMC
0.0083
Noise reference
T₁ 210 µs · CZ 0.2%
Published · 2025 Qiskit · Aer · QPE

Quantum principal component analysis for face recognition

Replaced the classical eigendecomposition at the heart of PCA face recognition with quantum phase estimation over a Trotterised Hamiltonian, applied to the normalised covariance density matrix. Eigenvalues and eigenfaces are recovered from the phase register and used directly for projection and recognition.

Evaluated on the Yale face set — 25 training images, 40 held-out probes, including impostor pairs to measure false-match and false-non-match rates across the decision threshold.

github.com/Rijuki/QPCA_Riga
Registers
5 phase + 5 data qubits
Subroutine
QPE + Suzuki–Trotter
Best correct prediction
39 / 40  (97.5%)
False match rate
0.0 at operating point
False non-match
0.025
Input basis
128 × 128 greyscale
Published · 2022 IBM Quantum · CHSH

Calibrating gate rotations against Bell's inequality

Running a CHSH test on hardware means choosing the rotation angles that maximise violation — a brute-force sweep that meant 100+ quantum jobs per candidate. We replaced it with a gradient-style index walk that perturbs, measures and steps, converging on the same optimum at a fraction of the job count.

Note: measurements above the theoretical 85% bound are noise, not quantum advantage. Calibrating against hardware means characterising the artefact.

github.com/Rijuki/BITgatepowers
Target
Tsirelson bound 85%
Measured
85 – 87%
Search
Index walk vs. full sweep
Cost
Same optimum, far fewer jobs
Method
CHSH / Bell's inequality
Reference
Hidary, Applied Approach

Also in the lab: Cu2N.QE — a density functional study of copper nitride with Quantum ESPRESSO, and an in-browser JupyterLite/WASM quantum kernel deployment. All repositories

649eras of walk-forward validation
97.5%recognition rate, quantum eigenspace
99.7%max orthogonality to classical baseline
156qubits on the reference Heron r2 processor
02

Where we consult

We take the engagements where the honest answer is still “it depends”.

i

Quantum ML feasibility & scoping

Which parts of your problem genuinely need a quantum circuit, and which are better served by a classical baseline you already own. Delivered as a written recommendation with a test plan.

ii

VQC architecture & training

Encoding strategy, ansatz depth, re-uploading schedule and an optimiser matched to your noise budget — from parameter-shift gradients to SPSA on hardware.

iii

Noise-aware deployment

Calibration extraction from live backends, noise injection during training, shot-count budgeting and post-measurement readout correction, sized to your actual QPU allocation.

iv

Benchmarking & honest evaluation

Era-based walk-forward validation, orthogonality measurement against your incumbent model, and reporting that distinguishes signal from noise. No leaderboard cherry-picking.

v

Quantum cryptography & post-quantum readiness

QKD protocol selection, key-management integration and migration planning against the incoming post-quantum standards landscape.

vi

Technical diligence for investors

Independent assessment of a quantum claim: is the advantage measured or simulated, does it survive noise, and is there a classical method that already wins?

03

How we work

Three commitments that shape every engagement.

01

Measure on hardware, early

Simulator results are a hypothesis. We pull the backend's calibration data in week one and train against it, so the architecture is shaped by the noise floor instead of discovering it at the end.

02

Report what didn't work

A negative result, properly measured, is a deliverable. Every project we publish carries its failed variants and their numbers alongside the headline.

03

Keep a classical control

Every quantum claim is stated relative to a strong classical baseline. If the baseline wins, we say so — that is the only way the quantum contribution can be trusted.

04

Start a conversation

Tell us the problem you're actually trying to solve. We'll tell you whether quantum belongs in it — including when the answer is no.

Email quantum@reneejulian.com
Office Barcelona, Spain
Code github.com/Rijuki