I'm a PhD researcher & MLOps Engineer at KTH Royal Institute of Technology (Stockholm) with 5+ years of end-to-end ML experience — from raw sensor data to production APIs monitored 24/7.

My core focus is building intelligent process monitoring systems that fuse multi-modal signals (vibration + acoustics) to predict surface quality in real-time manufacturing. I treat every model as a product: versioned, containerized, tested, and observable.

• 🔬 Research: Multimodal deep learning, LLMs for time-series, Vision Transformers, Diffusion Models for data augmentation
• ⚙️ Engineering: End-to-end MLOps (MLflow, Docker, Kubernetes, GitHub Actions, Evidently, Grafana)
• ☁️ Cloud: AWS (SageMaker, S3, Lambda, Kinesis, Step Functions), GCP, Azure Databricks
• 📦 Currently: Part-time ML Engineer @ NordML — demand forecasting system for Nudient AB on GCP
• 📍 Stockholm, Sweden

Specialties: LLMs (LoRA/QLoRA fine-tuning) · Vision Transformers (ViT) · CNNs (VGG, ResNet, InceptionResNet) · GANs · Diffusion Models · LSTM / RNN · NLP Transformers · SVM / RF / GBR

End-to-end intelligent process quality monitoring system

• Built a multimodal ML pipeline fusing vibration + acoustic signals to predict surface roughness in real time
• Achieved >15% error reduction vs. single-sensor baselines via ensemble fusion with attention mechanisms
• Productionized with MLflow + Docker + Kubernetes + GitHub Actions CI/CD; drift detection via Evidently; real-time dashboards in Grafana
• Deployed both batch and low-latency REST APIs; enforced data contracts for governed, auditable releases

Inventory demand forecasting system

• Built multi-SKU demand forecasting models with Shopify data pipelines, feature engineering, and DVC versioning
• Deployed batch + real-time inference on GCP with Docker + CI/CD
• Monitored via Grafana with MAPE, service level, and fill rate alerts tied to safety-stock replenishment

ML-based milling runout & tool life prediction

• Developed SVR, ANN, LR models on vibration time-series to predict tool wear impact from milling runout
• End-to-end responsibility: data collection → preprocessing → modeling → presentation at R&D monthly conference

Fine-tuned ChatGLM2 with LoRA/QLoRA for surface quality prediction from textualized vibration features. Improved cross-dataset accuracy by 10%; strong results in zero-shot and small-sample scenarios.

Applied Conditional GANs & Classifier-Guidance Diffusion Models to synthesize training images. 📄 Published in Procedia CIRP, 58th CIRP Conference on Manufacturing Systems

Modified & fine-tuned ViT with multiple attention mechanisms for multimodal surface quality prediction. 📄 Under review — Journal of Advanced Engineering Informatics

Leveraged pre-trained CNNs (VGG16/19, ResNet50V2, InceptionResNetV2) on audio data; hyperparameter tuning with Bayesian Optimization. 📄 Published in Journal of Manufacturing Letters

Industrial collaboration with Volvo Truck, Sandvik, Swerim AB — deployed models achieving 95%+ accuracy, R² > 0.98 for surface quality prediction. 📄 Published in Journal of Engineering Applications of Artificial Intelligence

Degree	Institution	Period
🎓 Ph.D. — Intelligent Manufacturing (ML Monitoring Systems)	KTH Royal Institute of Technology, Sweden	2019 – Present
🎓 M.Sc. — Production Engineering	KTH Royal Institute of Technology, Sweden	2015 – 2018
🎓 M.Sc. — Computer Science (Minor)	Aalto University, Finland	2015 – 2018
🎓 B.Sc. — Mechanical Engineering	Savonia University of Applied Sciences, Finland	2012 – 2015

💡 Setup note: Export the overview figure from each paper PDF → save as paper_images/paper1.png ... paper5.png in this repo to display the images below.

Leveraging classifier-guidance diffusion model for improved surface roughness prediction through synthesized audible sound signal

📰 Procedia CIRP, 58th CIRP Conference on Manufacturing Systems, 2025  |  🔗 DOI: 10.1016/j.procir.2025.XX.XXX

Applied a classifier-guided diffusion model to synthesize 2D Mel-spectrogram images of audible sound signals from machining, tackling data scarcity in deep learning-based monitoring. Synthetic + real data boosted VGG16 prediction accuracy and generalization on enriched datasets.

Key topics: Diffusion Model Data Augmentation Transfer Learning VGG16 Mel-spectrogram Surface Roughness

Surface quality prediction in-situ monitoring system: A deep transfer learning-based regression approach with audible signal

📰 Journal of Manufacturing Letters, 2024  |  🔗 DOI: 10.1016/j.mfglet.2024.09.156

Built an in-process surface roughness monitoring system using pre-trained CNNs (VGG16, VGG19, ResNet50V2, InceptionResNetV2) on Mel-spectrogram images of audible sound. Model architectures fine-tuned via Bayesian optimization; compared across multiple input window lengths.

Key topics: Transfer Learning CNN Bayesian Optimization Audible Sound Mel-spectrogram In-process Monitoring

Data-driven approaches for surface quality monitoring and prediction based on heterogeneous multi-channel signal fusion in hard part machining

📰 Journal of Engineering Applications of Artificial Intelligence, 2025  |  🔗 Open Access — CC BY

A comprehensive methodology evaluating signal selection, multi-sensor fusion, and ML model deployment for hard turning. Validated on two machining platforms; includes uncertainty quantification (NPKDE-based prediction intervals) and Bayesian hyperparameter tuning.

Key topics: Multi-sensor Fusion Hard Turning ML Models Uncertainty Quantification Bayesian Optimization Feature Engineering

Surface roughness monitoring and prediction based on audible sound signal with the comparison of statistical and automatic feature extraction methods in turning process

📰 Euspen's 24th International Conference & Exhibition, 2024  |  🔗 Open Access — CC BY

Benchmarked statistical (handcrafted) vs. automated feature extraction methods applied to audible sound signals for surface roughness prediction in CNC turning. Provided clear guidance on feature strategy selection for practical deployment.

Key topics: Feature Engineering Automated Feature Extraction Signal Processing Turning Process Time-series Analysis

A multi-modality transformer approach: surface roughness monitoring and prediction with bottleneck attention mechanisms

📰 Under Review — Journal of Advanced Engineering Informatics

Modified and fine-tuned Vision Transformer (ViT) models combined with multiple bottleneck attention mechanisms for surface roughness prediction under multimodal information fusion (vibration + acoustic). Benchmarked across multiple fusion scenarios.

Key topics: Vision Transformer (ViT) Attention Mechanisms Multimodal Fusion Deep Learning Surface Quality Prediction

Data augmenation based surface quality monitoring with machine learning models

📰 Oral presentation as the invited speaker at Digital Cluster Conference, Stockholm, 10-05-2023

 |  🎥 Video Link

Modified and fine-tuned Vision Transformer (ViT) models combined with multiple bottleneck attention mechanisms for surface roughness prediction under multimodal information fusion (vibration + acoustic). Benchmarked across multiple fusion scenarios.

Key topics: Data Augmentation Signal Synthesis ML Algorithms Surface Roughness Prediction Audible Sound