Phongsakon Mark Konrad

Flood detection using satellite imagery is crucial for environmental monitoring and disaster management, especially in rural and agricultural regions where even minor water inundation can disrupt farmland accessibility and road safety. Sentinel-1 Synthetic Aperture Radar (SAR) imagery offers a robust solution for mapping water under various weather conditions. Although deep learning-based segmentation methods have shown promising results for flood detection, their comparative performance in agricultural landscapes, including small-scale surface water dynamics, remains underexplored. In this study, we introduced a three-class segmentation framework that distinguishes sea, inland water, and land, improving the flood detection accuracy in complex coastal farmland. Ten different deep learning models were evaluated for segmentation using Sentinel-1 VH polarization decibel values. We further investigated anomaly detection via autoencoders and variational autoencoders to track temporal changes in flood-prone areas. The evaluations showed that the DeepLabv3+ and hybrid ResNet-UNet models outperformed the others.

Deep learning models are notoriously data-hungry, which presents a fundamental obstacle for specialized medical tasks where annotated data is scarce. We perform a carefully controlled experiment on a clinically important problem: segmenting the layers of the artery wall from only nine annotated histology images. We compare a variety of standard convolutional architectures, pre-trained on a large public histology corpus, against a vision foundation model guided by a systematic and reproducible prompting curriculum.

Gastrointestinal (GI) imaging modalities including endoscopy, colonoscopy, and wireless capsule endoscopy provide vital diagnostic information, yet their manual interpretation remains laborious. Although deep learning approaches, particularly convolutional neural networks (CNNs), hybrid architectures, and transformer-based models, have achieved high accuracy in this domain, their clinical adoption is impeded by data limitations and trust concerns. The current study systematically maps algorithmic trends to specific GI imaging techniques, quantifies the relationship between dataset size and model performance to identify data sufficiency thresholds, and evaluates translational enablers such as federated learning for data privacy and explainable AI for clinician trust.

Ad hoc dataset search requires matching underspecified natural-language queries against sparse, heterogeneous metadata records, a task where typical keyword or embedding retrieval alone falls short. We present a reference architecture for agentic hybrid retrieval that combines BM25 lexical search with dense-embedding retrieval via reciprocal rank fusion, orchestrated by a large language model controller that repeatedly plans queries, evaluates result sufficiency, and reranks candidates. To reduce the vocabulary mismatch between user intent and provider-authored metadata, we introduce an offline metadata augmentation step in which an LLM generates pseudo-queries and structured pseudo-descriptions for each dataset record.

Large language models are increasingly used as software architecture co-pilots, yet no benchmark evaluates their cloud-native architecture knowledge. We present CAKE, 188 expert-validated questions spanning four cognitive levels (recall, analyze, design, implement) and five cloud-native topics. We tested 22 model configurations (0.5B-70B parameters) from four families, using three-run majority voting for MCQ and LLM-as-judge scoring for free-response items. We observe four findings: (1) MCQ accuracy plateaus above 3B parameters, with 15 of 22 configurations exceeding 90%; (2) free-response scores scale steadily across all cognitive levels; (3) the two formats capture different facets of knowledge, as MCQ approaches a ceiling while free-response continues to differentiate models; (4) reasoning augmentation (+think) improves free-response quality, while tool augmentation (+tool) degrades performance for small models.

The current AI coding agents select frameworks, scaffold infrastructure, and wire up integrations. These are architectural decisions, yet almost no one reviews them as such. We study the problem from two angles. First, we survey agentic coding tools and identify five mechanisms by which they make implicit architectural choices. The case study shows that changing the prompt alone produces structurally different systems. Second, we analyze prompt-architecture coupling. The way prompt choices in LLM-integrated code dictate what infrastructure the system needs. Six recurring patterns arise from this analysis. Natural-language instructions now appear to shape system structure directly. We call this vibe architecting, and we outline review practices, decision records, and tooling aimed at bringing these hidden decisions under governance.

We propose a novel multi-layered AI framework for day-ahead electricity trading in European power markets, focusing on the Nordic region. Our framework combines three complementary components: (i) Machine Learning(ML)/Deep Learning(DL) forecasters modelling prices, load, weather and renewable generation; (ii) Reinforcement Learning (RL) agents choosing trading or bidding strategies based on these forecasts and world states (weather, current price, current load and current renewable generation); (iii) a layer of debating Large Language Models (LLMs) agents that debate about the input and output of previous layers, integrate textual news and produce final recommendation via an LLM judge. This design enables the system to 'see' through the forecasting models, 'act' through RL agent and 'explain' through language-based reasoning. We use publicly available data from Nord Pool, ENTSO-E and Energinet to evaluate each layer individually and in combination to improve both trading performance and interpretability.

Hyperspectral imaging enables non-destructive assessment of fruit ripeness and firmness, yet most published methods rely on deep learning architectures that are difficult to deploy in commercial sorting environments. This study benchmarks 19 traditional machine learning algorithms on dual-task prediction (ripeness classification and firmness estimation) using the DeepHS Fruit dataset with 1,120 engineered VIS-NIR spectral features across five fruit species. Through a 6-phase experimental design spanning 190 preprocessing configurations, 3,800 Bayesian optimization trials, and 10-fold cross-validation, we find that preprocessing strategy consistently matters more than algorithm choice: stratified resplit balancing and spectral feature engineering together account for the largest performance gains, while PCA dimensionality reduction degrades accuracy by 2.52 percentage points on average by discarding discriminative spectral information. The best single model (ExtraTrees with stratified resplit, no PCA) achieves 75.00% overall accuracy, surpassing Fruit-HSNet's reported benchmark of 70.73% while training in 0.31 seconds on consumer hardware. An RGB-only comparison reaches 69.57% accuracy (93% of full-spectrum performance), though this represents an upper bound from bands extracted from hyperspectral cubes rather than true RGB cameras. Explainable AI analysis via SHAP, LIME, and permutation importance reveals concentrated feature importance in visible regions associated with pigment changes and near-infrared regions related to tissue structure and water content, linking model decisions to known indicators of fruit maturation. These results demonstrate that carefully preprocessed traditional ML pipelines can match or exceed published benchmarks for spectral fruit quality tasks, offering a practical, interpretable, and resource-efficient alternative for commercial deployment.