Ivan Lorencin

The integration of natural language understanding into conversational commerce requires robustly extracting structured intents from idiomatic text. While cloud-based models are highly capable, their latency, privacy risks, and operational costs necessitate efficient local alternatives. To address this, we empirically evaluate locally deployed, open-weights Large Language Models (LLMs) on parsing multi-intent shopping cart commands into deterministic JSON schemas. We benchmark multiple models across minimal, extended, and few-shot prompting paradigms, utilizing a dual-axis framework to simultaneously measure extraction quality (operation F1-score, exact match, schema validity) and hardware performance (latency, GPU power consumption). Our evaluation highlights a fundamental trade-off between schema compliance and computational overhead, demonstrating how advanced prompting techniques improve parsing accuracy while measurably impacting inference latency and power draw. Ultimately, this comprehensive profiling identifies optimal model configurations for resource-constrained environments: a well-chosen 3-8B parameter model (specifically Granite4 3B) paired with few-shot prompting recovers over 95% of the quality of the largest evaluated model at a coarse indicative estimate of 40× lower operational energy per command, establishing a practical baseline for deploying cost-effective LLMs as efficiency candidates in resource-constrained environments.

This study evaluates the accuracy of medical-report interpretations generated by large language models in comparison to medically verified sources, with a particular focus on urology. The main goal is to examine the extent to which LLMs can reliably and precisely explain medical findings, with emphasis on expert urological terminology. For the evaluation of interpretation accuracy, real clinical documentation was used, specifically a dataset of 40 urological reports from the Clinical Hospital Center Rijeka. The generated explanations were compared with medically correct definitions using quantitative metrics assessing interpretability, specificity, and clinical accuracy. The results revealed a clear stratification of model performance, with DeepSeek-V3.2, GLM-4.6, MiMo-V2-Flash, and Qwen2.5-7B-Instruct achieving the highest semantic accuracy and consistency, while several models exhibited unstable behavior characterized by occasional catastrophic failures. Overall, the findings indicate that although LLMs can produce clinically coherent explanations of urological findings, their variability and susceptibility to hallucinations necessitate human oversight, supporting their use primarily as decision-support tools rather than autonomous clinical interpreters.

Recent advances in large language models (LLMs) highlight two dominant strategies for performance improvement: prompt engineering and fine-tuning. While prompt design can significantly influence model output, it remains uncertain whether lightweight fine-tuning methods, such as adapter-based training, offer meaningful advantages for structured, domain-specific tasks. This study builds on prior research comparing three prompting strategies for natural-language command parsing into JSON schemas. Expanding that framework, the current work investigates how adapter-based fine-tuning, where most model parameters are frozen and only small adapter modules are trained, affects model accuracy and consistency. The experiment uses the same controlled shopping-cart parsing task and dataset of 12,000 synthetic commands to ensure direct comparability. Results quantify the trade-off between computational cost and performance gains, offering evidence-based insights into whether fine-tuning is a justified investment compared to advanced prompt engineering. The contribution of this study is a clear, empirical framework for deciding when fine-tuning meaningfully enhances LLM utility in applied natural-language understanding.

Face recognition is increasingly being adopted in industries such as education, security, and personalized services. This research introduces a face recognition system that leverages the embedding capabilities of the CLIP model. The model is trained on multimodal data, such as images and text and it generates high-dimensional features, which are then stored in a vector index for further queries. The system is designed to facilitate accurate real-time identification, with potential applications in areas such as attendance tracking and security screening. Specific use cases include event check-ins, implementation of advanced security systems, and more. The process involves encoding known faces into high-dimensional vectors, indexing them using a vector index FAISS, and comparing them to unknown images based on L2 (euclidean) distance. Experimental results demonstrate a high accuracy that exceeds 90% and prove efficient scalability and good performance efficiency even in datasets with a high volume of entries. Notably, the system exhibits superior computational efficiency compared to traditional deep convolutional neural networks (CNNs), significantly reducing CPU load and memory consumption while maintaining competitive inference speeds. In the first iteration of experiments, the system achieved over 90% accuracy on live video feeds where each identity had a single reference video for both training and validation; however, when tested on a more challenging dataset with many low-quality classes, accuracy dropped to approximately 73%, highlighting the impact of dataset quality and variability on performance.

This paper presents a memetic algorithm (MA) for energy cost estimation of a robot path. The developed algorithm uses a random recombination genetic algorithm (GA) as the basis for the first stage of the algorithm and performs a local search based on feature importances determined from the data in the second stage. To allow for the faster determination of the solution quality, the algorithm uses an ML-driven fitness function, based on MLP, for the determination of path energy. The performed tests show that not only does the GA itself optimize the point-to-point paths well, but the usage of MA can lower the energy use by 58% on average (N = 100) when compared to a linear path between the same two points.

This study evaluates the performance of seven large language models (LLMs) in generating context-aware recommendations. The system is built on a collection of PDF documents (brochures) describing local events and activities, which are embedded into an FAISS vector store to support semantic retrieval. Synthetic user profiles are defined to simulate diverse preferences, while static weather conditions are incorporated to enhance the contextual relevance of recommendations. To further improve output quality, a reranking step, utilizing Cohere’s API, is used to refine the top retrieved results before passing them to the LLMs for final response generation. This allows better semantic organization of relevant content in line with user context. The main aim of this research is to identify which models best integrate multimodal inputs, such as user intent, profile attributes, environmental context and how these insights can inform the development of adaptive, personalized recommendation systems. The main contribution of this study is a structured comparative analysis of 7 LLMs, applied to a tourism-specific RAG framework, providing practical insights into how effectively different models integrate contextual factors to produce personalized recommendations. The evaluation revealed notable differences in model performance, with Qwen and Phi emerging as the strongest performers, whereas LLaMA frequently produced irrelevant recommendations. Moreover, many models favored gastronomy-related venues over other types of attractions. These findings indicate that although the RAG framework provides a solid foundation, the selection of underlying models plays an important role in achieving high quality recommendations.

Electronic health records (EHRs) are typically stored in relational databases, making them difficult to query for nontechnical users, especially under privacy constraints. We evaluate two practical clinical NLP workflows, natural language to SQL (NL2SQL) for EHR querying and retrieval-augmented generation for clinical question answering (RAG-QA), with a focus on privacy-preserving deployment. We benchmark nine large language models, spanning open-weight options (DeepSeek V3/V3.1, Llama-3.3-70B, Qwen2.5-32B, Mixtral-8 × 22B, BioMistral-7B, and GPT-OSS-20B) and proprietary APIs (GPT-4o and GPT-5). The models were chosen to represent a diverse cross-section spanning sparse MoE, dense general-purpose, domain-adapted, and proprietary LLMs. On MIMICSQL (27,000 generations; nine models × three runs), the best NL2SQL execution accuracy (EX) is 66.1% (GPT-4o), followed by 64.6% (GPT-5). Among open-weight models, DeepSeek V3.1 reaches 59.8% EX, while DeepSeek V3 reaches 58.8%, with Llama-3.3-70B at 54.5% and BioMistral-7B achieving only 11.8%, underscoring a persistent gap relative to general-domain benchmarks. We introduce SQL-EC, a deterministic SQL error-classification framework with adjudication, revealing string mismatches as the dominant failure (86.3%), followed by query-join misinterpretations (49.7%), while incorrect aggregation-function usage accounts for only 6.7%. This highlights lexical/ontology grounding as the key bottleneck for NL2SQL in the biomedical domain. For RAG-QA, evaluated on 100 synthetic patient records across 20 questions (54,000 reference–generation pairs; three runs), BLEU and ROUGE-L fluctuate more strongly across models, whereas BERTScore remains high on most, with DeepSeek V3.1 and GPT-4o among the top performers; pairwise t-tests confirm that significant differences were observed among the LLMs. Cost–performance analysis based on measured token usage shows per-query costs ranging from USD 0.000285 (GPT-OSS-20B) to USD 0.005918 (GPT-4o); DeepSeek V3.1 offers the best open-weight cost–accuracy trade-off, and GPT-5 provides a balanced API alternative. Overall, the privacy-conscious RAG-QA attains strong semantic fidelity, whereas the clinical NL2SQL remains brittle under lexical variation. SQL-EC pinpoints actionable failure modes, motivating ontology-aware normalization and schema-linked prompting for robust clinical querying.