Kaspars – Python, LLM, AI agent development

An AI-powered product discovery platform for construction materials, enabling users to search and explore large product catalogs using semantic text and image search. The platform featured hybrid RAG search, intelligent filtering, automated catalog ingestion, and computer vision models for product classification and color extraction. I led the technical architecture and development of the AI and search infrastructure.

• Engineered a hybrid search utilizing RAG, text, and image embeddings, and custom ranking for enhanced search capabilities.
• Architected a dynamic faceting system to streamline complex product taxonomies and improve catalog navigation for construction professionals.
• Developed automated ingestion pipelines with computer vision models for precise classification and color extraction of material catalogs.

A boutique software development company headquartered in Riga, Latvia. The agency specializes in building scalable backend infrastructure, custom content management systems, and specialized REST/JSON API integrations tailored for small and medium-sized businesses.

• Directed the development of an end-to-end AI application for transforming user prompts into structured short-form videos.
• Designed a comprehensive orchestration workflow covering script creation, scene planning, and asset generation.
• Implemented asynchronous execution and scene-level retries to enhance the application's reliability and efficiency.
• Trained and integrated YOLO-based defect detection models on Nvidia Jetson using TensorRT.

An AI platform for personalized print-on-demand products that automates design analysis, enhancement, and production preparation. The platform used computer vision, multimodal machine learning, LLMs, semantic search, and image generation models to classify designs, estimate embroidery costs, improve image quality, detect design properties, and retrieve visually similar assets.

• Applied LoRA and GRPO reinforcement learning to post-train a multimodal foundation model for production personalization workflows.
• Built training datasets, evaluation benchmarks, and post-training workflows for continuous model improvement.
• Created and deployed a product personalization pipeline using FastAPI, image upscaling, LLM reasoning, and Nano Banana Pro.
• Trained a design decomposition model using object detection, Pix2Pix, SPADE, and CycleGAN on 65K+ designs.
• Built a ResNet color profile classifier with 97% accuracy and about 3,000 daily predictions.
• Trained a multimodal model on 300K+ images and tabular data to estimate embroidery costs.
• Integrated a super-resolution model for image upscaling.
• Built a TensorFlow.js model for semi-transparency detection.
• Created a design similarity search PoC using FAISS, Elasticsearch, ResNet embeddings, perceptual hashing, and custom ranking.
• Used MLflow and Weights & Biases for experiment tracking, model versioning, and performance monitoring.

Enterprise data and backend solutions and contribution to AI-driven proof-of-concepts for a leading global professional services company that specializes in digital transformation, cloud computing, cybersecurity, and artificial intelligence.

• Developed an urban city traffic light controlling PoC with SUMO using reinforcement learning.
• Developed an agent-based intelligent control system.
• Researched algorithms.
• Trained models.
• Developed API backend functionality and unit tests.
• Developed new features, hot fixes, and new endpoints.
• Improved functionality in collaboration with co-located teams from India, the USA, and a data analytics team from Riga.