Machine Learning Engineer

Bluetweak is an omnichannel customer support platform that uses AI to improve workflows and response quality across communication channels.

• Developed and implemented an intelligent response system that combined semantic template matching with domain-specific Sentence Transformers, LLM-based template validation, and RAG knowledge retrieval.
• Fine-tuned the Stella400M model on real customer interactions to enhance semantic similarity matching and contextual understanding for domain-specific responses.
• Integrated hybrid response selection logic to ensure high accuracy while maintaining strict business communication standards.
• Improved customer support consistency by automating knowledge lookups and context retrieval from internal documentation.

Apsisware is a subsidiary of Arnia Software that provides comprehensive machine learning solutions for clients in eCommerce, food tech, and retail sectors.

• Developed and deployed object detection models (Mask R-CNN, YOLOv5) for a fashion recommender app and food recognition pipeline, achieving over 90% mAP.
• Created classification (ResNet) and segmentation (U-Net) models for automated product tagging and food item identification.
• Deployed optimized inference models on Raspberry Pi 4 devices using the NCNN C++ framework, including model quantization.
• Contributed to a 3D apartment reconstruction pipeline from monocular video utilizing Droid-SLAM, Polygon-Transformer, and Cube R-CNN.
• Built a shopping assistant with an in-domain BERT classifier (F1 score 0.93) and fine-tuned an Octopus 2B model with Q-LoRA for query routing (accuracy 0.85).
• Implemented a RAG-based recipe recommendation system for a grocery retailer using Llama Index.
• Developed an anomaly detection pipeline for identifying foreign objects in ovens through a hybrid VLM reasoning approach (Qwen2-VL + Llama3-8B).

October is a neo-lending platform that provides fast financing to SMEs, aiming to disrupt traditional banking with data-driven credit processes.

• Designed a PySpark-based feature engineering and model training pipeline from transactional datasets.
• Built and deployed an LGBM credit risk model that reduced SME loan default rates from ~11% to ~2%.
• Developed an OCR-based tool that processed financial statements from scanned PDFs using OpenCV, AWS Textract, and Lambda functions.
• Delivered APIs for real-time risk scoring and integrated them with internal credit approval workflows.

Software Improvement Group (SIG) is a Netherlands-based company specializing in software quality assurance and risk assessment.

• Developed an end-to-end machine learning pipeline for image-based defect detection on carton packaging.
• Applied image segmentation and classification models (Mask R-CNN, EfficientNet) to identify micro-defects and misprints in high-resolution scans.
• Created an automated labeling pipeline using weak supervision techniques to accelerate dataset creation.
• Integrated the solution into the production line’s quality control workflow, reducing manual inspection times by 60%.