04/14/2026 updated
100 % available
ML Engineer / Senior Software Engineer
Prishtine, Kosovo
Worldwide
Master of Science (M.Sc.)About me
Senior SWE turned ML Engineer. 12 yrs production systems + MSc ML from TU Munich. Built state-of-the-art multi-modal deep learning models on 44K+ MRI scans. Full ML lifecycle ownership. Ready to ship AI products.
Python (Programming Language)Node.jsTensorFlowTransformers (Electrical)PyTorchDeep LearningTerraformDocker
12 years building and leading production systems as a senior full-stack and cloud engineer (AWS, Kubernetes, Node.js, Python). MSc in Machine Learning from TU Munich, with research at the Lab for AI in Medicine — built deep learning models (CNNs, Vision Transformers, multi-modal architectures) for biological age prediction on 44,000+ MRI scans, achieving state-of-the-art results. Additional ML research experience in multi-modal video understanding, surpassing state-of-the-art on the HVU dataset. Currently building AI-powered products including a real-time medical transcription and clinical summary platform using Whisper, Claude, and AWS. Comfortable owning the full ML lifecycle: data pipelines, model training, experiment tracking, interpretability, and deployment.
Languages
GermanBasic knowledgeEnglishGood
Project history
Architected a multi-tenant SaaS platform for clinic and hospital management, serving multiple medical specialties across 6 languages (EN, DE, DE-CH, SQ, FR, IT).
Built a real-time audio transcription pipeline using WebSocket (Socket.IO) and AWS Transcribe Streaming, with automatic fallback to Whisper large-v3 (via Groq) for unsupported languages — including a custom hallucination filter using segment-level character ratio detection.
Implemented dynamic provider routing — database-driven configuration selecting between AWS Transcribe and Groq Whisper based on language and clinic type.
Designed an AI post-processing pipeline using Claude for medical terminology correction with specialty-aware prompts (ENT, cardiology, physiotherapy, dental), and automated clinical summary generation outputting structured notes in the tenant's configured language.
Full-stack delivery: NestJS backend, Next.js 14 web client, React Native mobile app, PostgreSQL/Prisma, all provisioned via Terraform on AWS.
Designed and trained CNN and Vision Transformer models (ResNet-50, DeiT-small, 3D ResNet-18) for biological age estimation from MRI data, working with 44,000+ participants from the UK Biobank dataset.
Proposed the first multi-modal age prediction model integrating whole-body and brain MRI, achieving a MAE of 1.83 years — outperforming single-modality baselines.
Compared post-hoc interpretability methods (Grad-CAM, Grad-CAM++) with intrinsic transformer interpretability (attention rollout), identifying cardiovascular region and knee joints as primary aging markers.
Implemented full training pipelines in PyTorch including transfer learning, data augmentation, Z-score normalization, and model checkpointing on NVIDIA A40 GPU.
Designed a novel multi-modal transformer model for multi-label video classification, fusing audio, video, speech, and optical flow modalities across 3,142 semantic categories.
Achieved state-of-the-art on the Holistic Video Understanding (HVU) dataset — 57% mAP, surpassing the previous best (FrameExit, 47.7%) despite training on only 90% of available data.
Benchmarked against TimeSformer, analyzing tradeoffs between temporal-only and divided space-time attention.
Built a parallelized data pipeline on AWS S3 to process and store 432,000+ training videos.