International Journal of Automation, Artificial Intelligence and Machine Learning

Breaking Down Barriers: How Conversational AI Makes Smart Homes Accessible to All

Praveen Chinnusamy — Wed, 15 Oct 2025 00:00:00 -0500

Objective: Smart home technology promises convenience and independence, yet it systematically excludes elderly users, people with disabilities and non-technical populations due to rigid command syntax, platform fragmentation and privacy concerns. We introduce SmartHomeHarmonizer, an edge-based conversational framework that enables natural language interaction across heterogeneous voice platforms while preserving user privacy.

Methods: Our framework combines a progressive neural architecture compression pipeline, a dynamic cross-platform intent mapping protocol and a federated accent-adaptation mechanism. These innovations reduce model size from 438 MB to 28.7 MB with negligible accuracy loss, unify Alexa, Google Assistant, HomeKit and Matter ecosystems through semantic bridging (a technique that maps natural language intents to platform-specific APIs using ontological relationships), and continuously adapt to user accents on-device using differential privacy. We evaluate SmartHomeHarmonizer in three six-month deployments (assisted living facilities N=142, disability services N=48 and affordable housing N=89), using Bonferroni-corrected statistical analysis to compare against existing edge and cloud assistants.

Results: SmartHomeHarmonizer achieves 89.3% intent recognition accuracy, reduces task completion time by 68.7% (p<0.001, Cohen's d=1.82) and lowers end-to-end latency to 584 ms, outperforming state-of-the-art edge solutions. Adoption reached 82% among elderly users, while caregiver assistance requests dropped by 54%. Federated learning improved recognition accuracy for non-native speakers from 71.3% to 91.6% over six months without compromising privacy. The system operates at 3.1 W idle power consumption, representing a 73% reduction compared to commercial alternatives.

Conclusion: Our results demonstrate that sophisticated conversational AI can run efficiently on low-cost hardware (<US$55) and significantly improve accessibility. By releasing the source code, datasets, trained models, and deployment infrastructure, SmartHomeHarmonizer provides a replicable foundation for inclusive smart-home research.

Role of Quality Assurance in DevOps: Bridging the Gap Between Development and Operations

Wumi Ajayi, Idowu Olugbenga Adewumi, Nelson Ayibawanemi John — Mon, 10 Nov 2025 00:00:00 -0600

The way we incorporate Quality Assurance (QA) into DevOps has come a long way. It is no longer just a step that happens after development; now, it is a continuous, smart process that flows throughout the entire software delivery lifecycle. This study explores how blending AI and machine learning techniques with modern QA practices can make a real difference, drawing on data from various production environments. The results are impressive: we found that defect density dropped by as much as 64.2%, the average time to repair issues fell by 42.1%, automated test coverage increased by 39.1%, and deployment success rates went up by 16.6%. By employing model-driven strategies like predicting build failures with XGBoost, generating intelligent test cases using CodeT5, and detecting anomalies through Isolation Forest, our integrated framework is able to spot risks early, enhance test execution, and speed up the release process. When we compared these outcomes to our benchmarks before the integration, it became clear that AI-enhanced QA not only reduces production defects and rollback incidents, but it also helps eliminate the bottlenecks typical of traditional QA methods. These study revealed the game-changing potential of self-healing, predictive QA systems for handling scalable, high-frequency release cycles. Nevertheless, it is important to note some trade-offs, such as the added complexity of maintaining test suites and the impact on pipeline execution time.

A Hybrid Cloud-Based AI Framework for Real-Time Insurance Fraud Detection Using Snowflake and Cortex AI

Sreenivasarao Amirineni — Tue, 18 Nov 2025 00:00:00 -0600

Insurance fraud remains one of the most pervasive challenges in the financial sector, demanding scalable and explainable AI-driven solutions. This study presents an integrated framework that combines snowflake’s cloud-native data warehouse and ML Ops architecture with cortex AI’s machine-learning environment to enable real-time detection of fraudulent insurance claims. The proposed system unifies supervised algorithms (XGBoost, logistic regression) and unsupervised techniques (isolation forest, LOF) to balance high-accuracy classification with anomaly discovery. Experimental evaluation on a dataset exceeding one million claims achieved 88% accuracy and an F1-score of 0.87, confirming the framework’s robustness and scalability. A theoretical analysis explains the superior performance of supervised models due to the bias–variance balance and label-driven optimization, while exploratory deep-learning benchmarks (CNNs, LSTMs) illustrate potential gains in sequential fraud pattern detection. The inclusion of Snowflake’s model registry, lineage tracking, and multi-cloud integration ensures compliance and operational transparency. Future extensions involving autoencoders, GANs, and federated learning are proposed to advance adaptive, cross-institutional fraud analytics. Overall, the hybrid snowflake–cortex AI architecture represents a reproducible, enterprise-ready solution for modernizing fraud detection with transparency, efficiency, and regulatory alignment.

From Reactive BI to Agentic AI: The Rise of Domain Specific Autonomous Insight Systems in Enterprise Analytics

Praveen Koushik Satyanarayana — Wed, 19 Nov 2025 04:37:12 -0600

Enterprises need descriptive and diagnostic insight engines that think ahead of users, not after them. This paper presents a research framework for agentic analytics that unifies four ideas into a single, verifiable system. First, ontology aligned grounding maps business language to concrete data assets and permitted joins. Second, a scaffolded tool layer executes plans safely across query engines, catalogs, code runners, and visualization services while exposing traces for audit. Third, a multi-turn hypothesis engine treats explanations as first-class objects that are generated, tested against data, revised, and ranked. Fourth, a dual judge mechanism fuses an LLM critic with a gold data layer that serves as the final arbiter of numerical claims and structural correctness. We specify rubric signals for each stage and show how those signals become rewards for reinforcement learning of both single agents and coordinated teams. The result is a pathway from natural language intent to governed, reproducible descriptive analytics that improve over time.

Artificial Intelligence and Control Charts: A Big Problem

Fausto Galetto — Wed, 17 Dec 2025 04:12:01 -0600

We use the data of some published papers to compare those authors findings with ours. From the analysis we get different results: the cause is that the authors use the probability limits of the PI (Probability Interval) as they were the confidence limits (control limits of the control charts, CCs). The control limits in the Shewhart CCs are based on the normal distribution Central Limit Theorem (CLT) and are not valid for non-normal distributed data: consequently, the decisions about the “In Control” (IC) and “Out of Control” (OOC) states of the process are wrong. The control limits of the CCs are wrongly computed, due to unsound knowledge of the fundamental concept of confidence interval. Minitab and other software e (e.g. JMP, SAS) use the “T charts”, claimed to be a good method for dealing with “rare events”, but their computed control limits of the CCs are wrong. The same happens for the confidence limits of the parameters of the distribution involved in the papers (Weibull, Inverse Weibull, Gamma, Binomial, Maxwell). We will show that the Reliability Integral Theory (RIT) is able to solve these problems and the Sequential way of dealing with data.