Case Studies : Production ML for Swiss Enterprise

Case Studies

where research meets production

Richemont

Aesthetical Quality Control

Computer Vision · Luxury Manufacturing · Neuchâtel, Switzerland

Automating aesthetical and dimensional quality inspection for luxury manufacturing. The challenge: industrial-grade accuracy meeting luxury-grade standards, where every fraction of a pixel matters and false positives are as costly as missed defects.

Applied to your business

If you run visual quality control on a production line, this same subpixel defect detection can be tuned to your tolerances and deployed on standard hardware.

Read full case study →

Mission 01

Subpixel Image Alignment

High-precision image registration for multi-gigapixel imagery, achieving subpixel accuracy on production captures.

OpenCVNumPyCUDANative extensionsMemory optimization

Outcomes

Sub-pixel accuracy on production imagery

Multi-gigapixel processing on standard hardware

View details

Developed and benchmarked advanced image alignment algorithms to register high-resolution captures of watch components with subpixel precision.

Techniques

Subpixel registrationHomography manipulationsTransformersRandom sampling algorithms

Mission 02

Feature Engineering & Dimensionality Reduction

Advanced feature extraction and manifold visualization for high-dimensional quality metrics, enabling interpretable decision boundaries and robust feature selection pipelines.

scikit-learnNumPyPythonManifold visualization

Outcomes

Interpretable quality feature spaces

Robust feature selection pipeline

View details

Researched and implemented specialized dimensionality reduction techniques to transform high-dimensional image data into discriminative, compact feature representations. Explored manifold learning approaches tailored to the specific structure of surface inspection data.

Techniques

Unsupervised methodsManifold learningManifold visualizationCustom feature extraction

Mission 03

ML & Deep Learning Anomaly Detection

Production-grade anomaly detection combining statistical methods with deep learning architectures for automated aesthetical quality assessment at scale.

PyTorchscikit-learnStatistical anomaly detectorsDeep learning architectures

Outcomes

Production-grade detection accuracy

Automated quality classification

View details

Built a comprehensive anomaly detection pipeline that combines classical machine learning approaches with deep learning architectures for identifying aesthetical defects with high recall and controlled false-positive rates.

Techniques

Statistical anomaly detectorsSOTA papers implementationsDeep learning architecturesEnsemble methods

Mission 04

Depth Estimation & Surface Properties

Depth estimation, normal, texture, and shape reconstruction for surface inspection beyond 2D appearance.

PyTorchOpenCVNeural surface modelingDepth estimation methods

Outcomes

Surface geometry object enhancing

Exploratory research and SOTA methods combined

View details

Explored depth estimation and surface-reconstruction methods.

Techniques

Reconstruction methodsTexture and Shape methods Several relighting methods

Mission 05

Software Architecture & Packaging

Modular pipeline architecture with vector similarity search, optimized memory management, and real-time processing throughput on standard industrial hardware.

FastAPIDockerPythonRelational databasesVector search librariesCUDA

Outcomes

Modular, configuration-driven architecture

Real-time processing throughput

View details

Designed the overall software architecture for the quality control system from hardware image acquisition through processing stages to result storage and visualization. Built robust data pipelines that orchestrate multi-stage workflows, handle error recovery, and maintain data lineage across the full inspection chain. The architecture supports modular algorithm swapping and configuration-driven pipeline composition.

Techniques

Pipeline orchestrationConfiguration-driven designError recoveryData lineage trackingModular architecture

Mission 06

MLOps & Deployment

Containerized production deployment with CI/CD pipelines, monitoring, and web-based dashboards for operator and stakeholder reporting.

DockerCI/CD pipelinesWeb-based dashboardsMonitoring

Outcomes

Deployed in production

Iterative, versioned release cycles

Selected Projects

Selected Engagements

Kering

Supply Chain Replenishment Optimization

Supply-Chain ML · Luxury Retail

PythonReinforcement LearningStatistical ModelingSimulationPandasNumPyPythonReinforcement LearningStatistical ModelingSimulationPandasNumPy

Read full case study →

Challenge

Kering's luxury retail warehouses faced a classic inventory dilemma: minimize average stock levels to reduce carrying costs while keeping missing sales below acceptable thresholds. Static replenishment rules could not adapt to seasonal demand shifts and promotional events, leading to either excess inventory or costly stockouts.

Approach

Run as a capstone collaboration with Kering, the work explored and developed a dynamic replenishment algorithm combining statistical demand modeling with reinforcement learning.

Reduced average stock levelsControlled missing sales rateValidated via simulation backtesting

How we'd apply this

If inventory cost and stockouts pull against each other, the same demand-modelling and simulation approach can size dynamic replenishment for your network.

Bracco Suisse SA

Liver Lesion Classification from Contrast-Enhanced Ultrasound

Medical Imaging AI · Healthcare

PyTorchMedical ImagingCEUS ProcessingPythonDeep LearningPyTorchMedical ImagingCEUS ProcessingPythonDeep Learning

Read full case study →

Challenge

Bracco Suisse needed to improve diagnostic confidence in contrast-enhanced ultrasound examinations by automatically classifying liver lesion types from noisy, low-contrast CEUS images where temporal contrast dynamics and image artifacts make standard classification approaches unreliable.

Approach

Designed custom neural network architectures for CEUS image processing, implementing domain-specific augmentation strategies that account for ultrasound artifacts, contrast agent dynamics, and temporal perfusion patterns. Built a preprocessing pipeline tailored to the unique characteristics of contrast-enhanced ultrasound data.

Multi-class lesion classificationCEUS-specific preprocessing pipelineDomain-adapted augmentation strategies

How we'd apply this

If you work with a noisy imaging modality, we build the domain-specific preprocessing and classification pipelines that off-the-shelf models miss.

from industry to academia

Specialized Research

Research Initiatives

01Applied ML Research · Bioinformatics

Protein Functional Classification

Have sequence and structured data to classify? We benchmark graph, sequence, and fusion models to find what genuinely fits your data.

PyTorchGNNCNN+2

Read case study →

02Computer Vision Research · Life Sciences

Embryo Survival Prediction from Time-Lapse Video

Need predictions from video or time-lapse imagery? We build temporal pipelines that read change over time, not just single frames.

PyTorchCNNLSTM+2

Read case study →

03Streaming Data & ML · Fintech

Real-Time Stock Market Prediction

Need decisions on live data? We design ingestion and online-learning pipelines that score events the moment they arrive.

KafkaHadoopSpark+3

Read case study →

04Explainable ML · Digital Health

Parkinson's Detection via Gait Analysis

Need a model a domain expert will actually trust? We pair strong predictors with SHAP explainability built for decision support.

XGBoostRandom ForestVAE+2

Read case study →

01Applied ML Research · Bioinformatics

Protein Functional Classification

Have sequence and structured data to classify? We benchmark graph, sequence, and fusion models to find what genuinely fits your data.

PyTorchGNNCNN+2

Read case study →

02Computer Vision Research · Life Sciences

Embryo Survival Prediction from Time-Lapse Video

Need predictions from video or time-lapse imagery? We build temporal pipelines that read change over time, not just single frames.

PyTorchCNNLSTM+2

Read case study →

03Streaming Data & ML · Fintech

Real-Time Stock Market Prediction

Need decisions on live data? We design ingestion and online-learning pipelines that score events the moment they arrive.

KafkaHadoopSpark+3

Read case study →

04Explainable ML · Digital Health

Parkinson's Detection via Gait Analysis

Need a model a domain expert will actually trust? We pair strong predictors with SHAP explainability built for decision support.

XGBoostRandom ForestVAE+2

Read case study →

Let's Talk

Let's discuss your next project

Every engagement starts with a conversation. Tell us about your challenge.

DISCUSS YOUR PROJECT Explore Services