U.S. Bancorp

Online streaming video understanding requires models to process continuous visual inputs and respond to user queries in real time, where the unbounded stream and unpredictable query timing turn memory management into a central challenge. Existing methods typically compress visual tokens via visual similarity heuristics, or augment compression with KV-cache-level retrieval. However, compression decisions rarely incorporate semantic signals, and retrieval is often added after compression is finalized, making the two stages hard to coordinate. We present SAVEMem, a training-free dual-stage framework that brings semantic awareness into memory generation and lets the retrieval scope adapt per query. In Stage 1, SAVEMem builds a three-tier streaming memory online under a constant memory budget. A fixed pseudo-question bank provides a lightweight semantic prior, so that long-term retention is shaped by semantic salience rather than visual similarity alone. In Stage 2, SAVEMem performs query-aware retrieval over this memory. An anchor-conditioned recency gate adapts the retrieval scope from short-term to mid- and long-term memory based on whether the query targets the present or the distant past. Within this scope, late interaction between query and memory tokens selects candidate frames for answering. Applied to Qwen2.5-VL without training, SAVEMem improves the OVO-Bench overall score from 52.27 to 62.69 and yields consistent gains on StreamingBench and ODV-Bench, while reducing peak GPU memory by 48% at 128 frames over the backbone.

The financial services industry faces mounting pressures to deliver real-time, personalized services while safeguarding sensitive user data under tight regulatory environments. Yet, prevailing AI systems in FinTech remain largely cloud dependent, which introduces latency bottlenecks, privacy exposure, and compliance risk. Meanwhile, industry analyses suggest that Edge AI is rapidly becoming a foundational shift, with predictions that 60% of AI deployments will run partially on device by 2029. However, existing edge AI research often focuses on inference optimization, not full-stack orchestration of financial microservices, and therefore, lacks the integrated, decision-oriented intelligence that is required to operate wholly on the device. In this work, we present an architecture for on-device microservice orchestration of generative AI tailored for FinTech use cases. Our system modularizes AI tasks, such as local LLM inference, fraud detection, biometric authentication, and credit scoring, into services coordinated via lightweight orchestrators (e.g. WASMEdge, Open Horizon). Unlike prior approaches, our system coordinates these services using lightweight WebAssembly-based runtimes, enabling secure, isolated, and efficient execution even on resource-constrained devices. Sensitive data, such as transaction history and biometric templates, remains strictly local, with optional federated synchronization for global fraud pattern sharing. With quantized LLMs, we attain inference latency under 90ms, while local anomaly detection achieves 72% accuracy in simulated financial fraud scenarios. The architecture integrates modular microservices, privacy-first orchestration, and a hybrid federated intelligence layer and is among the first to present a decentralized, compliant, and performance-sensitive AI infrastructure for the FinTech of reality.

The successive increase in the heterogeneous data in structured databases, semi-structured documents, and unstructured text repositories have rendered context-conscious knowledge retrieval a significant concern in the modern information systems. Sometimes, traditional methods of search such as keywords, rules, and other systems do not have the capability of reflecting the intent of the semantics, dependencies between contexts and cross-source relations leading to a common output of fragmented and low-precision retrieval. In an effort to overcome these shortcomings, the paper at hand suggests an LLM-based cognitive search model that would be used to achieve context-based knowledge retrieval in heterogeneous data sources. The suggested method consists of the combination of large language models (LLMs) and semantic embedding algorithms, contextual reasoning, and adaptive query orchestration to allow the deep interpretation of the user intent and data semantics. This structure integrates multi-modal and multi-source data with a semantic indexing layer and uses dynamically refined and ranked queries and synthesized knowledge with LLM-based reasoning. The proposed model in contrast to traditional retrieval systems enables contextual memory, conversational interaction, and domain-aware inference, which enable it to provide accurate, explainable and personalized search results. Experimental analysis on the variety of enterprise-scale datasets proves that the retrieval accuracy, contextual relevance, and response consistency are improved significantly in comparison with traditional and neural information retrieval benchmarks. The findings suggest that there is a next-generation of knowledge management, decision support system and enterprise information retrieval system which involves the LLM-enabled cognitive search as a scalable and intelligent solution.

Because IT infrastructure is growing in complexity and risk, disaster recovery (DR) is becoming increasingly difficult. Use of neuro-symbolic reasoning and quantum inspired optimization in a self-sufficient AI framework to help in operational resilience. Neuro-symbolic reasoning combines logic rules with deep learning methods. This can help devise strategies that promptly respond to their context in case of failure. The optimization of quantum inspiration is used for implementing the best recovery path. It is possible with resource allocations in big IT systems. Since lesser involvement of a human, adapt with the threat, and recovers more efficiently. Using what happens before, the system updates its symbolic rules and neural models with the knowledge-driven approach. The recovery process will utilize the available resources to minimize disruption and the framework will provide explainable decision making for compliance and audit. The simulation outcomes indicate a significant decrease in downtime and a significant increase in recovery time when compared to traditional disaster recovery solutions. It is a scalable and flexible offering for cloud-native environments and a multi-site IT infrastructure capable solution.

The fast proliferation of federated multi-agent systems in fields like smart healthcare, autonomous systems, edge-cloud intelligence, and cyber-physical systems has further raised the issue of ethical responsibility, openness, and regulation in the decentralized artificial intelligence (AI). In contrast to centralized AI systems, federated multi-agent systems use autonomous agents to learn and operate together in heterogeneous and untrusted environments, and traditional governance mechanisms are insufficient. The paper suggests a federation-specific Ethical AI Governance Model, and it is based on the principles of fairness, explainable, privacy protection, accountability, and regulatory compliance. The suggested model proposes a multi-layer governance architecture to include the ethical policy orchestration, compliance enforcement at the agent level, safe federated auditing, and ongoing assessment of ethical risk with the help of trust-aware metrics. The framework also guarantees accountable decision-making by integrating governance controls both in the local agent and global federation levels without affecting the system scalability and learning efficiency. The widespread conceptual analysis and illustrative examples reveal the effectiveness of the proposed model of governance in the reduction of ethical risks, including the spread of bias, the lack of transparency in decision making, and the development of malicious agents. The research provides the groundwork of federated intelligence based on ethics-by-design, which allows the reliance on reliable, transparent, and socially responsible implementation of the large-scale multi-agent AI systems.