An efficient water supply management system can be one of the applications of the Internet of Things (IoT). Water is a basic physiological need, and smart management of water plays a significant role in a smart city. This paper focuses on a mathematical model and IoT framework that aid in developing a smart city. A framework is developed for water supply management. The efficiency of the water supply can be measured by monitoring leakage conditions, overflow of water, automatic meter reading and online bill payments, and water consumption status of households, community, state, and eventually the whole country as well as the automatic water supply line cut-off. The system where the IoT is being deployed consists of embedded hardware in which sensors and microcontrollers provide messages and gain feedback from each other with the help of the internet, and this process can not only be monitored but also can be controlled from a remote location. The developed framework addresses all these aspects and mathematical equations are used and formulated while developing the IoT application. The mathematical equations are concentrated on consumption level (CL), leakage reporting (LR), and bill amount (BA) based on consumption. These become the point of contact for deploying IoT and eventually a framework is developed. This framework can be useful not only in water supply management but also in the management of road traffic, pollution, garbage, home automation and so on. In a nutshell, this paper illustrates the usage of the IoT framework in water supply management which contributes to developing the smart city.

Video surveillance is widely used in different areas like roads, malls, education, industries, retail, parks, bus stands, and restaurants, each presenting distinct anomaly patterns that demand specialized detection strategies. Adapting anomaly detection models to new camera viewpoints or environmental variations within the same scenario remains a significant challenge. Extending these models to entirely different surveillance environments or scenarios often requires extensive retraining, which can be both resource-intensive and time-consuming. To overcome these limitations, model frameworks, i.e., the video anomaly detector model, have been proposed, leveraging the meta-learning framework for faster adaptation using swin transformer for feature extraction to new concepts. In response, the dataset named MSAD (multi-scenario anomaly detection) having 14 different scenarios from multiple camera views, is the high resolution anomaly detection dataset that includes diverse motion patterns and challenging variations such as varying lighting and weather conditions, offering a robust foundation for training advanced anomaly detection models. Experiments validate the effectiveness of the proposed framework, which integrates model-agnostic meta-learning (MAML) with a ten-shot, one-query adaptation strategy. Leveraging the swin transformer as a spatial feature extractor, the model captures rich hierarchical representations from surveillance videos. This combination enables rapid generalization to novel viewpoints within the same scenario and maintains competitive performance when deployed in entirely new environments. These results highlight the strength of MAML in few-shot learning settings and demonstrate its potential for scalable anomaly detection across diverse surveillance scenarios.

Video surveillance is widely used in different area like road, mall, education, industries, retail, park, Bus stand, Restaurant and smart cities, each presenting unique anomalies requiring specialized detection. However, adapting anomaly detection models to novel viewpoints within the same scenario poses challenges. Extending these models to entirely new scenarios necessitate retraining or fine-tuning, a process that can be time-consuming. To address these challenges, Model named Video Anomaly Detector Model has been proposed, leveraging the meta learning framework for faster adaptation using Swin Transformer for feature extraction to new concepts. In response, The Multi-Scenario Anomaly Detection (MSAD) dataset, featuring 14 scenarios from various camera views, is the first high-resolution anomaly detection dataset for real-world applications. It includes diverse motion patterns and challenging variations like different lighting and weather conditions, providing a strong foundation for training advanced models. Experiments confirm the effectiveness of MAML, particularly when trained on the MSAD dataset. MAML excels under new viewpoints within the same scenario and performs competitively in entirely new scenarios, showcasing its potential for detecting anomalies in diverse and evolving surveillance environments.

Dahal Jeet is a four-player, fixed-partnership, trick-taking card game played daily across the Terai-Madhesh region of Nepal, in which only the four tens score. Its objective is lexicographic and non-additive: a four-valued primary criterion with a trick-count tiebreak that activates on the 2–2 split and decides a third of hands, where the trick-taking games used as benchmarks all score additively. It belongs to the documented Mendikot / Dehla Pakad branch of the Court Piece family, matching Mendikot on 21 of 26 compared features, but the Nepali variant itself has no published rule description: the standard reference’s survey of card games played in Nepal lists no game of this branch, and searches of OpenAlex, arXiv, Shodhganga and Nepal Journals Online (NepJOL), in English and Devanagari, return one paper on the entire family. This paper contributes, first, a formal specification of the ruleset drawn from long-term insider participant observation across seven districts and two generations of players; and second, a computational analysis of the game built on that specification. We implement and compare 26 agents spanning hand-written heuristics, metaheuristic tuning, determinized search (PIMC, single- and multiple-observer ISMCTS), and seven families of learning agent, evaluated under duplicate-deal protocols with bootstrap confidence intervals and family-wise error control over 325 pairwise comparisons. Search dominates: the strongest agent is ISMCTS (Elo 1663.2), and no learning method exceeds a tuned heuristic. An exact double-dummy solver provides an endgame optimality bound, showing search converts 0.923 of theoretically won endgames against 0.810 for the best heuristic. We further report a negative result of practical interest: a deliberately minimal state encoding outperforms a rich hand-designed one by +0.0644 win rate, with no seed overlap. Every deterministic heuristic tested is exploitable by a search agent that knows its policy, while random play is not, and a variance decomposition attributes 29–39% of outcome variance to the deal itself. Conditioning on the deal exposes the result the title names: in a game where only the tens score, trump length predicts the outcome far more strongly than the tens dealt, moving the win rate by 0.592 against 0.254. Finally, a human trial with 10 players from 8 districts, 20 hands each, shows why the deal must be controlled before a win rate means anything: the strongest agent’s raw rate against them, 0.4800 (95% CI [0.4200, 0.5350], resampling participants), is indistinguishable from chance, yet replaying each participant’s own cards with a stand-in agent in their seat gives a paired deficit for the human of-0.0930 (95% CI [-0.1450,-0.0380]). The agent’s advantage over people is real and small, and invisible to the unpaired comparison. The protocol is not specific to this game: wherever chance dominates outcome variance, replaying each participant’s own realised deal with a stand-in agent recovers a comparison that the unpaired analysis reverses

Large multilingual decoder language models publish closed lists of supported languages: Llama 3.1 and Aya-23 both claim Hindi but not its lower-resource Devanagari sisters, Maithili, Nepali, and Bhojpuri. We ask a mechanistic question: where, and how far, can a single inference-time direction move such a model toward an unclaimed sister language, using only its Hindi representations and no fine-tuning, no new data, and no weight updates? Because the four languages share the Devanagari script and most subword tokens, anyshift mustreflect representation geometry rather than vocabulary. Across two architecturally distinct eight-billion-parameter decoders we report four findings. First, both models exhibit a shared hourglass geometry (languages distinct at the surface, merged mid network, and re-separated near the output), yet language identity is linearly decodable at every layer (about 99.5% accuracy). Second, adding the difference of per-language activation centroids at the re-divergence onset shifts generation toward the target (peak language adherence 4.03 of 5 for Nepali on Llama, scored by a three-judge cross-family panel) but only at the surface: fluency collapses into repetition, and no combination of model, target, and coefficient is simultaneously target-adherent and fluent. Third, few-shot prompting does reach fluent sisters, so the ceiling is a limit of single-direction steering, not model competence; steerability tracks linear separability, not proximity. Fourth, Hindi outperforms English as the steering source despite English’s pretraining dominance. Steering thus serves as a causal probe of where language identity is manipulable. We release a four-way parallel Devanagari corpus and an evaluation harness.