Research Systems
Applied AI systems built at the intersection of safety, governance, and real-world deployment — each tied to peer-reviewed research.
Featured Projects
LagrangianCTDE — Constrained MARL for Disaster Response
Risk-aware multi-agent reinforcement learning framework coordinating Storm, Flood, and Evacuation agents under Lagrangian safety constraints, achieving 81.5 reward with only 2.3% safety violations across six baselines.
Problem: Coordinating heterogeneous agents in disaster scenarios often results in unsafe behaviors where performance is prioritized over civilian safety or resource preservation.
Technical Approach: Implemented a Constrained Centralized Training with Decentralized Execution (CTDE) architecture using Lagrangian multipliers to enforce strict safety bounds on agent actions during training.
Agentic AI-Enhanced Digital Twin — Smart City Infrastructure
Physics-grounded simulation framework evaluating rule-based, digital twin, and agentic AI monitoring architectures for smart city civil infrastructure, with blockchain-anchored audit trails and Kalman-filtered state estimation.
Problem: Traditional infrastructure monitoring relies on static thresholds that fail to detect complex structural decay or sensor spoofing attacks.
Technical Approach: Developed a multi-agent hierarchy where 'Observer' agents perform Kalman filtering and 'Auditor' agents log state transitions to a Hyperledger-based blockchain for immutable verification.
FinNutriAgent — Household Budget & Nutrition Optimizer
Open agentic AI framework jointly optimizing household meal planning and financial budgets under nutritional, cultural, and economic constraints using MILP and LLM orchestration across multi-store price data.
Prompt Injection Defense Framework
LangChain prompt-injection defense framework with a drop-in middleware, three detection modes, four policy strategies, a formal threat model, and a reproducible evaluation suite for blocking or annotating malicious prompts and tool content.
Problem: Agentic LLM pipelines are vulnerable to indirect prompt injection where untrusted tool outputs manipulate the orchestrator into performing unauthorized actions.
Technical Approach: Built a modular defense layer utilizing semantic analysis and rule-based sanitization to isolate system instructions from data-driven content.
TV-FLIDS — Trust-Aware & Verifiable Federated Intrusion Detection System
Byzantine-resilient federated learning system for IoT intrusion detection with a three-criteria verification gate, dynamic trust scoring, adaptive aggregation weights, and reproducible NSL-KDD evaluation pipelines.
All Projects
PAAI — Privacy-Aware Agentic AI for IoT Healthcare
Four-layer privacy-preserving multi-agent architecture for personalized chronic disease management, combining PPO reinforcement learning with AES-256 encryption, hash-chain audit logging, and 3-tier Human-in-the-Loop governance.
EAGF — Ethical AI Governance for Renewable Energy IoT
Four-pillar governance framework integrating transparency, fairness, privacy, and accountability for AI-driven cybersecurity in renewable energy IoT systems, validated with a Composite Trust Index across biometric and 5G solar-microgrid domains.
Research Literature Review Agent
Automated agentic pipeline that retrieves, individually summarizes, and synthesizes arXiv papers into publication-quality literature reviews with LaTeX export, powered by LangChain, GPT-4o, and FAISS semantic search.
TrustGuard — Autonomous Mobile Permission Governance
Multi-agent reinforcement learning framework for autonomous mobile permission governance using Constrained MAPPO with Lagrangian safety bounds, achieving 96.3% AUROC and 41.3% privacy risk reduction with only 2.1% false-revocation rate.
XAI Systematic Review 2026
PRISMA 2020-compliant systematic review synthesizing 188 peer-reviewed studies on Explainable AI across healthcare, finance, cybersecurity, robotics, and agentic AI domains, identifying SHAP and LIME as dominant techniques with critical coverage gaps.
IoUT Interrogator Framework
Interrogator-based behavioral trust inference framework for Internet of Underwater Things networks, using transformer temporal modeling, metadata-driven monitoring, and continuous trust scoring for privacy-preserving anomaly detection.
Smart Application Intelligence System
Mobile-first Flutter and FastAPI platform for tracking student applications, scoring fit and risk, ranking opportunities, and integrating SOP analysis with transparent decision logic and local-first data persistence.
AAIRM — Agentic AI Inventory Replenishment and Management
Research framework for multi-category retail inventory optimization that unifies forecasting, replenishment optimization, supplier-aware execution, and governance checks with reproducible benchmark pipelines.
Wildfire Governance-Constrained Agentic AI
Governance-Invariant MDP framework for safety-critical wildfire monitoring with blockchain-enforced Human-in-the-Loop oversight, achieving policy-agnostic safety guarantees and robust adversarial performance.
TRAIL NLP — Trust-Aware Cross-Lingual Alignment
Low-resource NLP framework that jointly optimizes cross-lingual alignment, probabilistic calibration, and entropy regularization to improve reliability and trust calibration for multilingual LLM inference.
ADAPT — Agentic AI Nutrition & Healthcare Monitor
Production-ready multi-agent PRA system for inclusive nutrition and healthcare support, coordinating meal planning, reminders, food guidance, and monitoring through a blackboard-driven orchestrator with XAI explanations, policy checks, and a FastAPI API.
AquaAgent — Smart Water Distribution Leak Detection
Proactive, policy-aware multi-agent system for real-time leak detection and governance in urban water distribution networks, built around an EPANET 2.2 digital twin with monitoring, anomaly detection, decision, and governance agents.
Privacy-Preserving Federated Learning Against Inference Attacks in Sensor Data
Federated learning pipeline for evaluating privacy-utility tradeoffs under honest-but-curious server attacks on sensor data, with configurable defenses, attack models, and reproducible HAR-based experiments.
EFADT-Smart-Campus — Explainable Federated Agentic Digital Twins
An explainable federated agentic digital twin framework designed for real-time smart campus resource optimization, combining privacy-preserving federated learning, XAI interpretability, and multi-agent coordination.
Wearable Physiological Monitoring & Analysis Package
Study replication and machine learning package for wearable-based continuous physiological monitoring in women's health, incorporating time-series sensor processing, feature extraction, and anomaly detection.
Wildfire-RL — Reinforcement Learning for Aerial Wildfire Suppression
Simulation environment and deep reinforcement learning benchmark for evaluating autonomous aerial suppression and evacuation routing policies under stochastic wind and thermal dynamics.
Secure Agent Runtime — Sandboxed Autonomous Execution Engine
Policy-enforced execution engine for autonomous AI agents, establishing strict action boundaries, capability isolation, runtime permission checks, and cryptographically signed action logs.
Automated Canary Deployment Coordinator
Automated canary deployment orchestration framework for microservices and AI model endpoints, featuring real-time health telemetry analysis, dynamic traffic shifting, and automatic rollback triggers.