Household energy consumption is strongly influenced by user habits and behaviors, making behavior-aware energy services essential to improve efficiency in smart environments. The Internet of Behaviors (IoB) has emerged as a promising approach that combines real-time Internet of Things (IoT) data, behavioral analytics, and intelligent decision services to support sustainable consumption practices. In this study, we propose a data-driven service engineering approach to analyze household energy use in the recent Residential Energy Consumption Survey (RECS 2020) dataset. We design a scalable analytics pipeline to extract distinct behavioral archetypes, understand energy consumption patterns, and explore how various user behaviors affect energy consumption. By applying data analytics for feature engineering and selection, followed by machine learning for clustering, our analysis reveals the primary drivers of energy consumption and identifies different user consumption profiles. Our findings offer valuable information for developing targeted strategies to optimize energy use in smart homes, providing a comprehensive understanding of the behavioral factors that support sustainable energy solutions. The study highlights the role of AI-enabled behavioral services as a key enabler for sustainable smart energy systems and contributes methodological insights to the engineering of next-generation data-driven software services.
Retrieval-Augmented Generation (RAG) grounds large language model (LLM) outputs in external knowledge but remains fundamentally static: once deployed, conventional RAG cannot learn from user feedback. This limitation is acute in resource-constrained deployments where smaller models require continuous adaptation. In such LLM-based services, adapter weights, retrieval indices, and prompt templates form an implicit behavioral layer lacking lifecycle management. We propose a general architectural pattern—automated feedback-to-training conversion, parameter-efficient adaptation, and zero-downtime deployment—instantiated as FitRAG via counterfactual training signal generation, LoRA fine-tuning, and a hot-swappable adapter architecture. On a 20-query quantum-computing QA benchmark, using response length as a diagnostic signal across five ablation studies, RAG retrieval is associated with 15.7% shorter responses, LoRA fine-tuning with 11.0% longer domain-specific outputs, and the feedback loop with 18.1% more concise responses; the full pipeline achieves ROUGE-1 0.171 and a same-family LLM-as-judge score of 0.80. These results suggest that feedback-driven continuous learning over the implicit behavioral layer is a viable direction for self-improving LLM-based services in resource-constrained settings.
Generative artificial intelligence (GenAI) data governance faces a key vulnerability: ensuring creator consent and fair royalties through accurate sources and trusted identity associations. With existing systems failing under digital modifications, we built a powerful blockchain network and proposed TrustGen, a decentralized service for scalable GenAI data management. TrustGen replaces fragile visual hashing with a decoupled two-stage decentralized digital authentication architecture to verify creator identity proofs.To overcome smart contract memory limitations, TrustGen extracts DINOv2 [CLS] tokens and applies QR-based orthogonal PCA to compress features into an EVM-compatible 256-dimensional space. This suppresses hardware-induced numerical drift and minimizes on-chain storage. Integrated XFeat eliminates unreliable matching and achieves sub-pixel alignment.To build a decentralized digital identity modeling framework based on the C2PA standard, creator metadata is stored on IPFS. Simultaneously, zero-intrusion smart contracts automatically allocate royalties, achieving trusted on-chain and off-chain identity collaboration.TrustGen achieves 8.0x latency acceleration and maintains stability under strong interference (Recall@10 ≥ 96%), while reducing Ethereum gas costs by 91.8%. Stress tests confirm a peak throughput of 1,701.5 QPS with only 9.4 ms latency. Processing 10 million records, TrustGen meets real-time SLAs, forming a highly feasible system component for enhancing on-chain trust.
AI coding agents can solve a large fraction of real-world software engineering tasks, yet the complexity and maintainability of their patches remain poorly understood and largely uncontrolled. We present Wily v2, a high-performance code-analysis tool that computes code complexity at over 570000 LOC/s, and integrate it as a complexity backpressure signal inside an agentic coding harness. To evaluate Wily on substantive tasks, we construct SWE-bench Complex, a dataset of 116 real GitHub issues from major Python projects with a median patch size of 48 lines (6.9× SWE-bench Verified). We evaluated three frontier models under control and Wily-feedback conditions. A single round of Wily feedback significantly reduces cyclomatic complexity by 20–31% and cognitive complexity by 39–54% across all three models.
