Research Notes

Overview

As smart cities transition toward autonomous urban commerce, automated procurement and inventory management systems are granted operational authority over high-frequency supply decisions. However, unconstrained autonomy introduces grave risks of untracked inventory drift, supplier disputes, and data sovereignty violations. Published in Frontiers in Artificial Intelligence, this research presents AAIRM (Governance-Aware Autonomous Retail Coordination), a framework addressing three requirements for trustworthy procurement: tamper-evident decision provenance, data sovereignty, and adaptive demand coordination. AAIRM couples a multi-agent Proximal Policy Optimization (PPO) ordering policy, a permissioned Blockchain Trust Ledger (BTL) for multi-party audit trails, and a Federated Demand Learning (FDL) layer. Evaluated on a multi-category synthetic environment and the public M5 dataset, AAIRM lowers normalized inventory cost by 13.2% (95% CI 12.2–14.2) against an ROP–EOQ baseline (10.2% on M5) at its cost-optimal operating point (92.3% fill), maintaining 8.2-point and 4.9-point advantages at 96.1% and 97.4% fill along its own cost–service frontier.

Key Contributions

• ›Rigorous Empirical Performance: Lowers normalized inventory cost by 13.2% (95% CI 12.2–14.2) against a conventionally parameterized ROP–EOQ baseline, and by 10.2% (95% CI 8.0–12.4) on the public M5 retail dataset at its cost-optimal operating point (92.3% fill)
• ›Layered Ablation Decomposition: Isolates 8.8 points of advantage to the reinforcement learning policy and 3.8 points to the coordination-and-governance block (feasibility projection, escalation logic, cross-category policy rules, supplier ranking, and negotiation)
• ›Critical Evaluation of LM Orchestration: Shows the language-model orchestrator contributes only 0.6 points (95% CI −0.9 to 2.1), statistically indistinguishable from zero, proving the defensible advantages are hard feasibility and auditability, not LLM reasoning
• ›Tamper-Evident Ledger with Minimal Overhead: The Blockchain Trust Ledger (BTL) adds just 6.6% decision-cycle overhead, validated through 500-case mutation replay and an explicit adversary model analyzing colluding majority limits
• ›Decentralized FedProx Demand Learning: Converges within 0.9 WAPE points of centralized training at a 1.3-point cost penalty, preserving retail data sovereignty across distributed stores
• ›Reward-Hacking Discovery & Mitigation: A 500-SKU evaluation exposes an insidious reward-hacking failure in high-spoilage categories, successfully neutralized through category-specific governance constraints

Practical Implications

Bridges the gap between autonomous capability and institutional accountability in AI cities. By rigorously dissecting where algorithmic advantages actually come from — showing that hard mathematical feasibility and cryptographic auditability drive value while language models add negligible benefit — the work provides regulators and urban planners with concrete, actionable governance instruments for overseeing autonomous commerce.

Overview

As connected vehicles move toward 5G/6G edge intelligence, federated learning (FL) is promoted as a privacy-preserving framework for collaborative model training across vehicles and roadside infrastructure without centralizing raw telemetry. However, gradient updates transmitted to the aggregation server retain distinct behavioral signatures. Presented at the IEEE 104th Vehicular Technology Conference (VTC2026-Fall), this research systematically investigates server-side client identity inference from weight deltas using inertial measurement unit (IMU) telemetry on the UCI HAR benchmark as an accessible proxy for onboard connected vehicle streams. Across five attack classifiers and five non-IID client partitions, an honest-but-curious server recovers client identities with near-perfect accuracy (≈1.000) from undefended updates. To neutralize this vulnerability, the paper evaluates a lightweight clip-then-noise defense swept across Gaussian noise levels under fixed clipping (C = 1.0) with formal (ε, δ)-differential privacy accounted via Rényi DP, alongside an Ensemble FL architecture providing a 1/K structural anonymity-set bound. The findings reveal a practical operating region (σ ∈ [0.1, 0.2]) that suppresses attack accuracy to near-random while incurring less than 5% relative utility loss.

Key Contributions

• ›Empirical demonstration of identity leakage: Proves that undefended weight deltas in federated learning leak sufficient behavioral patterns for an honest-but-curious server to achieve ≈1.000 client identification accuracy across five non-IID partitions and five attack classifiers
• ›Clip-then-noise defense with Rényi DP accounting: Evaluates Gaussian noise perturbations (σ ∈ [0.00, 1.00]) under fixed L2 clipping (C = 1.0), identifying a Pareto-optimal operating sweet spot (σ ∈ [0.1, 0.2]) that drives attacker advantage to near-random while preserving >95% relative model utility
• ›Ensemble FL structural privacy: Introduces an ensemble partitioning strategy that enforces a 1/K anonymity-set bound without adding noise, providing complementary defense against server-side de-anonymization
• ›Cryptographically verifiable evaluation: Guarantees rigorous evaluation through SHA-256 gradient disjointness verification between train and evaluation sets, count-normalized attacker-advantage metrics, and evaluation across multiple random seeds
• ›Open-source research pipeline: Released full reproducible implementation and evaluation scripts on GitHub (aliakarma/PPFL-Sensors)

Practical Implications

Confirms that privacy cannot be taken for granted in vehicular edge AI simply because raw sensor data is kept local. In safety-critical V2X environments where vehicle location, driving habits, and identities must remain untraceable, federated systems require principled differential-privacy noise injection or structural ensemble aggregation to protect user anonymity against curious server operators.

Overview

Assistive technologies for individuals with disabilities frequently suffer from fragmentation, lack of contextual reasoning, and rigid interfaces. Published in the IGI Global book 'Improving Disability Support and Institutional Performance with AI', this chapter proposes a comprehensive Agentic AI framework designed to deliver inclusive, autonomous, and personalized assistive ecosystems. By coordinating specialized agents for routine scheduling, health and nutrition support, and daily task management under a unified Perception–Reasoning–Action loop, the system dynamically adapts to individual accessibility requirements. A robust governance layer incorporates privacy-preserving policies, human-in-the-loop oversight, and auditable verification to ensure ethical, safe, and trustworthy operation.

Key Contributions

• ›Multi-agent architecture coordinating autonomous specialized agents for personalized daily routine management, health support, and accessibility accommodations
• ›Perception–Reasoning–Action (PRA) loop integrating multimodal inputs and adaptive contextual reasoning
• ›Governance-first design incorporating privacy-preserving data pipelines, human-in-the-loop oversight, and auditable verification
• ›Scalable framework bridging the gap between autonomous AI capabilities and safe, inclusive real-world deployment for people with disabilities

Practical Implications

Demonstrates how agentic AI architectures can move beyond isolated assistance toward holistic, governance-conscious ecosystems that enhance autonomy, reduce caregiver fatigue, and uphold dignity and equity for people with disabilities.

Overview

Mobile robots working alongside people and critical facilities need to reach their goals cheaply despite not knowing true traversal costs in advance and despite imperfect actuation. Solving the underlying stochastic shortest path problem exactly with value iteration costs computation that grows with the diameter of the map, while Dijkstra's algorithm is fast but conventionally considered inexact under stochastic transitions. This work characterizes exactly when that conventional wisdom is wrong: Dijkstra remains an exact planning engine whenever a reduced cost defined on the determinized map is nonnegative — a condition much weaker than the causality condition usually assumed. Building on that characterization, DORA (Dijkstra Oracle Reduced-cost Algorithm) learns online by calling a shortest-path oracle a fixed number of times per episode, never estimating a transition kernel, and adding a logarithmic survival weight when the probability of contact with a dynamic obstacle has to stay inside a budget.

Key Contributions

• ›A weaker exactness condition for Dijkstra under stochastic transitions: nonnegativity of a reduced cost on the determinized map, replacing the stronger causality condition commonly invoked
• ›DORA, an online learner that calls a shortest-path oracle a fixed number of times per episode and never estimates a transition kernel
• ›A logarithmic survival weight that keeps the probability of contact with a dynamic obstacle within a specified budget
• ›Evaluation across grid world navigation, directional drilling, and drone surveillance: matches optimistic value iteration given the true transition kernel while doing 4.5 to 19.3 times less planner work, and cuts contacts during learning by a factor of seventeen relative to determinize-and-replan

Practical Implications

Planner work, not model accuracy, is often the binding constraint for robots that must replan online. Because the exactness condition is checkable on the determinized map and the oracle is called a fixed number of times per episode, ordinary shortest-path search — already fast, well understood, and widely implemented — becomes a defensible basis for safe online navigation, with contact rates held inside budgets spanning two orders of magnitude.

Overview

pytest, JUnit, and Jest made it cheap to write low-level tests and run them on every commit, while systems engineering developed rigorous verification and validation principles for aligning requirements with deliverables. In practice the two worlds rarely meet, and for AI-enabled and cyber-physical systems that gap is increasingly costly: regulators now ask for traceable evidence that high-level requirements are met, but raw test results carry none of the structure such evidence requires. VNVSpec is an open-source framework in which users state high-level requirements directly or import them from catalogs derived from published standards, decompose them into module-level requirements with explicit metrics and acceptance criteria, and link them to test results through a traceability graph that compiles into verdicts and audit-ready reports.

Key Contributions

• ›Machine-readable, executable V&V specifications linking high-level requirements to the low-level tests already running in CI
• ›Requirement quality checking plus decomposition into module-level requirements with explicit metrics and acceptance criteria
• ›Traceability graph that compiles collected test evidence into verdicts and audit-ready reports
• ›Self-application evaluation: continuously assessed in CI against its own specification of 36 requirements verified by 449 tests, scaling linearly to roughly 10,000 requirements

Practical Implications

Compliance evidence stops being a document written after the fact and becomes a by-product of the test suite a team already runs. The same traceability machinery extends to black-box AI models and AI coding agents, where 'did this system meet the stated requirement?' is exactly the question that raw pass/fail output cannot answer.

Overview

Deploying AI anomaly detection in 5G solar-microgrids and industrial IoT presents severe governance challenges: black-box opacity, bias against rare protocol classes, and vulnerability to inference attacks. EAGF maps the EU AI Act's core requirements into engineering metrics governed across the full lifecycle. Fairness and privacy are co-optimized via Pareto multi-objective optimization, transparency is enforced via clarity-triggered pruning, and accountability is logged post-hoc to calculate a unified Composite Trust Index.

Key Contributions

• ›Mapping of EU AI Act pillars (Transparency, Fairness, Privacy, Accountability) to quantitative engineering metrics
• ›Pareto-guided multi-objective co-optimization of domain-adaptive fairness loss and privacy bounds
• ›Clarity-triggered structural model pruning for transparent inference
• ›Validated on biometric and Edge-IIoTset benchmarks, achieving +69.3% Trust Index gain with only +0.2 ms latency overhead

Practical Implications

Proves that regulatory compliance (EU AI Act) can be built directly into the AI training-and-deployment lifecycle of critical infrastructure with negligible operational overhead.

Overview

Traditional smart-city digital twins are passive simulators. Embedding agentic AI creates a proactive planning and control system where autonomous agents continuously test interventions in the simulated city before deploying them in the real one—dramatically reducing the risk of large-scale system changes while enabling optimization at scales no human team could manage.

Key Contributions

• ›Agentic AI integration architecture for smart-city digital twin platforms
• ›Autonomous optimization loops for civil infrastructure, traffic management, and emergency services
• ›Real-time simulation-driven policy testing before physical deployment
• ›Blockchain-anchored provenance log ensuring auditability of municipal interventions

Practical Implications

As megacities face intensifying pressure from population growth and climate change, AI-enhanced digital twins represent the most scalable approach to urban management. The simulation-first policy testing paradigm also provides an accountability layer for autonomous urban decisions.

Overview

Individuals with neurodivergent conditions or chronic disabilities often face cognitive overload when managing complex health routines and dietary needs. ADAPT introduces a Perception–Reasoning–Action multi-agent framework comprising meal planning, adaptive reminder, food guidance, and physiological monitoring agents. Operating via an MCP routing layer and shared blackboard, ADAPT combines hybrid rule-RL reasoning with XAI explainability to provide sensory-sensitive, personalized assistance.

Key Contributions

• ›Four-agent architecture (Meal Planning, Reminders, Food Guidance, Physiological Monitoring) in a PRA loop
• ›Model Context Protocol (MCP) routing layer paired with shared blackboard memory
• ›Hybrid reasoning combining rule-based safety constraints with reinforcement learning personalization
• ›Explainable AI (XAI) module providing sensory-sensitive, caregiver-verifiable explanations

Practical Implications

Offers a standardized, accessible blueprint for building personal health management systems that cater directly to neurodiverse populations, bridging the gap between clinical data and daily assistive technology.

Overview

Centralizing sensitive disability records across institutions violates privacy regulations, preventing the development of fair AI job-matching models. FedAgent-Chain resolves this by linking institutions in Saudi Arabia, the US, China, and Europe through differential-privacy federated learning (ε=3.2, δ=10⁻⁵). Rule-based agentic services and human-in-the-loop governance manage individualized accommodation pathways while a permissioned blockchain records audit trails.

Key Contributions

• ›Cross-institutional federated learning architecture with (ε=3.2, δ=10⁻⁵) differential privacy guarantees
• ›Multi-agent education-to-employment pathway generation with multilingual support
• ›Multi-dimensional fairness regularization reducing disability-category disparity by 32.3% and work-mode disparity by 40.3%
• ›Blockchain-anchored provenance log for accountable and audit-ready institutional decisions

Practical Implications

Demonstrates how privacy-preserving federated learning combined with agentic governance can solve sensitive cross-border workforce inclusion challenges without compromising data sovereignty.

Overview

Single-objective optimization is straightforward; multi-objective optimization with real-world constraints is hard. FinNutriAgent frames household food planning as a constrained multi-objective problem, using preference-based alignment to weight budget and nutrition goals according to user priorities and dietary requirements—producing Pareto-optimal weekly plans that adapt dynamically to price changes.

Key Contributions

• ›Dual-objective constrained reasoning pipeline for simultaneous budget and nutrition optimization
• ›Preference-based alignment enabling user-defined trade-offs between cost and nutritional goals
• ›Dynamic adaptation to ingredient price fluctuations without full replanning
• ›Evaluation against single-objective baselines demonstrating consistent Pareto improvements

Practical Implications

Multi-objective agentic planning has wide applicability—from personalized healthcare regimen planning to resource allocation in humanitarian aid. FinNutriAgent demonstrates that LLM-based agents can handle formal constraint satisfaction when properly scaffolded.

Overview

Urbanization and climate change demand intelligent infrastructure capable of proactive climate adaptation. Following PRISMA guidelines across 896 records (2018–2026), this study reviews 60 eligible papers (14 meeting the 4-property agentic threshold) for SDG 11 and SDG 13. The paper presents a reference architecture linking an agentic layer with an urban digital twin and conducts a real-data feasibility study on the SEVIR archive testing multimodal hazard classification.

Key Contributions

• ›PRISMA-guided rapid review of 896 records (60 eligible, 14 agentic) for urban climate resilience (SDG 11 & 13)
• ›Operationalized 4-property criteria (task autonomy, goal planning, tool use, multi-agent coordination) defining Agentic AI
• ›Reference architecture specifying state exchange between an agentic orchestration layer and an urban digital twin
• ›Real-data feasibility study on the SEVIR archive testing multimodal fusion against persistence baselines

Practical Implications

Provides municipal leaders and urban AI researchers with a comprehensive roadmap for moving from fragmented, domain-specific AI models to holistic, multi-agent climate resilience platforms.

Overview

Urban transport networks require dynamic, real-time traffic signal adjustments that adapt to non-stationary congestion patterns while maintaining physical safety. This paper introduces an agentic AI-driven traffic optimization architecture integrated with a real-time digital twin. Edge sensors feed physical telemetry into a perception layer; a LangChain conceptualization layer uses LLM reasoning and rule constraints to evaluate signal plans; and an action layer invokes Model Context Protocol (MCP) tool bindings to safely update physical traffic controllers.

Key Contributions

• ›Three-layer Perception–Conceptualization–Action framework for autonomous traffic signal control
• ›Model Context Protocol (MCP) integration for secure and standardized physical API execution
• ›Real-time digital twin co-simulation evaluating candidate timing plans before deployment
• ›Demonstrated reduction in intersection wait times and queue lengths over fixed-time and RL baselines

Practical Implications

Integrating LLM reasoning with MCP tool calls and digital twin verification provides a blueprint for safe autonomous control of municipal infrastructure, ensuring that high-level policies translate into deterministic, low-latency physical actions.

Overview

The proliferation of large language models in academic work raises urgent questions about authorship, originality, skill development, and equity. This paper rejects both uncritical adoption and reflexive prohibition in favor of a principled framework that specifies permitted uses, required disclosures, and prohibited applications across learning contexts—from undergraduate coursework to doctoral research.

Key Contributions

• ›Taxonomy of AI tool use cases by academic context and integrity risk level
• ›Disclosure and attribution standards for AI-assisted academic work
• ›Equity analysis of AI tool access across student populations
• ›Institutional policy adoption roadmap with pilot assessment protocols

Practical Implications

Without institutional guidance, AI tool use in academia will stratify rather than democratize access to academic success. Clear, fair guidelines benefit students, faculty, and administrators equally—and position academic institutions as responsible stewards of the AI transition rather than passive bystanders.

Overview

Current AI-based wildfire monitoring systems lack adaptive inter-agent coordination, structured human control, and cryptographically verifiable accountability. This work addresses those gaps by modeling wildfire monitoring as a constrained Partially Observable Markov Decision Process (POMDP) with explicit governance constraints baked into the reward and state-transition structure. A permissioned blockchain layer enforces human-authorization as a smart-contract invariant before any alert is disseminated, providing non-repudiation and Byzantine-fault resilience.

Key Contributions

• ›Governance-constrained POMDP formulation for disaster-monitoring agents
• ›Hierarchical multi-agent UAV coordination with risk-adaptive policy reallocation
• ›Smart-contract-enforced human-in-the-loop authorization as a hard state-transition invariant
• ›Formal proofs of alert integrity, human control, and non-repudiation under Byzantine fault assumptions
• ›Security analysis showing resistance to alert injection, replay, and tampering attacks

Practical Implications

For emergency management agencies, this framework offers a principled path to deploying autonomous aerial monitoring that regulators and the public can trust. The model's design principle—accountability embedded in the control loop rather than bolted on afterwards—is generalizable to any safety-critical autonomous system.

Overview

Autonomous underwater vehicles and IoUT sensor nodes face a critical vulnerability: once authenticated, agents are implicitly trusted for the entire mission. A compromised or drifting node can silently corrupt sensor fusion or mission-critical routing. The Interrogator-Based Secure Framework addresses this by treating trust as a continuous, behavior-derived quantity rather than a binary credential.

Key Contributions

• ›Interrogator module using a lightweight transformer to derive dynamic trust scores from communication metadata
• ›Proportional monitoring and conditional mission-access restrictions for suspicious agents
• ›Permissioned blockchain consortium for tamper-proof, decentralized trust-evidence storage
• ›21.7% detection accuracy improvement over static-trust baselines with minimal energy overhead

Practical Implications

As underwater IoT deployments expand to environmental monitoring, infrastructure inspection, and defense applications, zero-trust architectures like this become essential. The low-overhead transformer approach is also applicable to resource-constrained terrestrial IoT and satellite swarm networks.

Overview

In middle-income households, the tension between food affordability and nutritional adequacy is constant. FinAgent tackles both objectives simultaneously through a modular multi-agent architecture where specialized agents for budgeting, nutrition, price monitoring, and health personalization share a knowledge base and communicate via a substitution graph to maintain nutrient quality at minimum cost.

Key Contributions

• ›Price-aware multi-agent architecture with four specialized agent roles
• ›Substitution graph mechanism for maintaining nutrition quality during food-price shocks
• ›12–18% cost reduction versus static weekly menus while exceeding 95% nutrient adequacy
• ›Resilience to 20–30% food-price volatility through real-time adaptation

Practical Implications

AI-driven nutrition-finance optimization has direct applications for food-bank meal planning, hospital dietary services, and social welfare programs in price-volatile economies—aligning with the UN SDG goals of Zero Hunger and Good Health.

Overview

Agentic AI systems in high-stakes domains face a fundamental accountability gap: autonomous reasoning and action pipelines operate faster than human auditors can follow. The Blockchain-Monitored Architecture maps the perception–reasoning–action cycle onto a permissioned blockchain governance layer that verifies inputs, evaluates recommended actions against policy, and immutably records execution outcomes.

Key Contributions

• ›End-to-end architecture linking LangChain agent reasoning to Hyperledger Fabric governance contracts
• ›MCP-integrated action executors with verifiable pre-/post-condition enforcement
• ›Experimental validation across smart inventory management, traffic-signal control, and healthcare monitoring
• ›Demonstrated prevention of unauthorized actions with full traceability and acceptable latency

Practical Implications

The framework provides a universal blueprint for deploying agentic AI in regulated industries—from hospital automation to smart grid management—where regulators require cryptographic proof of policy compliance at every decision step.

Overview

Static reorder-point inventory systems fail when demand patterns shift rapidly or supply becomes uncertain. The proposed agentic pipeline replaces static rules with reasoning agents that continuously observe inventory levels, forecast demand trends, and autonomously trigger procurement decisions—with configurable human-oversight checkpoints for high-value orders.

Key Contributions

• ›Agentic AI pipeline integrating demand forecasting and autonomous procurement
• ›Dynamic adaptation to supply uncertainty through multi-source supplier reasoning
• ›Configurable human-oversight integration for regulatory compliance
• ›Demonstrated reduction in stockouts and overstock situations

Practical Implications

For manufacturers and retailers facing increasingly volatile supply chains, autonomous replenishment agents can dramatically reduce carrying costs while improving service levels—a compelling case for agentic AI in operations management.

Overview

Most AI assistants are designed for neurotypical users, creating significant barriers for those with disabilities or neurodivergent conditions. This framework deploys adaptive interface agents alongside specialized agents for healthy eating guidance, daily routine structuring, and emotional support—all governed by user-defined preferences and caregiver oversight protocols.

Key Contributions

• ›Multi-agent architecture with adaptive UI agents, nutrition agents, and routine-structuring agents
• ›Personalized assistance frameworks informed by disability-specific research
• ›Inclusive design principles embedded at the architectural level rather than as post-hoc accommodations
• ›Validation across diverse disability profiles and neurodivergent conditions

Practical Implications

As AI assistants proliferate, inclusive design from the ground up—rather than retrofitted accessibility features—can meaningfully improve quality of life for hundreds of millions of people worldwide. This framework is a template for human-centered AI that takes neurodiversity seriously.

Overview

Cloudbursts strike fast. Conventional systems forecast rainfall, then separately trigger response protocols, creating dangerous latency. This agentic framework treats the entire event lifecycle—from atmospheric sensing through hydrological modeling to communication and routing—as one continuous, adaptive reasoning loop where each agent's output is the next agent's input in real time.

Key Contributions

• ›Closed-loop integration of sensing, forecasting, downscaling, hydrological modeling, and coordinated response
• ›Demonstrated improvements in forecast reliability, critical success index, and warning lead time
• ›Communication and routing agents that maximize population reach while minimizing evacuation errors
• ›Adaptive recalibration with transparent auditability for post-event analysis

Practical Implications

Integrated sensing-to-response pipelines are the future of climate resilience infrastructure, especially in mountainous regions like northern Pakistan where urban expansion increases cloudburst risk. The framework is extensible to any rapid-onset natural disaster.