Authors: Prudvi Saisaran Ponduru
Abstract: Recent advances in foundation models, multimodal learning, reasoning-oriented large language models, agentic workflows, and edge AI have expanded the capabilities of artificial intelligence systems. However, practical decision-making remains brittle because many systems optimize prediction quality while under-modeling intervention effects, uncertainty, safety constraints, latency budgets, and human accountability. This paper introduces AEGIS-DM, an adaptive, edge-aware, governed, interventional, and safe decision-making framework designed for AI systems deployed across emerging technology settings including agentic assistants, cyber-physical systems, healthcare decision support, and enterprise automation. The framework combines five layers: multimodal state representation, predictive scoring, causal effect estimation, simulator- or planner-based long-horizon optimization, and a governance layer for calibration, fairness, policy checks, logging, and human override. We further propose a cross-domain evaluation protocol using public resources such as Adult, D4RL, WebShop, ALFWorld, MIMIC-IV Demo, NASA CMAPSS, and M5, together with open-source tooling including OpenAI Evals, Responsible AI Toolbox, OpenSpiel, RecSim NG, Stable-Baselines3, and RLlib. Because this manuscript is a methods-and-benchmark contribution, the quantitative section reports deterministic scenario-based simulation results under the stated protocol rather than production deployment measurements. Under the reference protocol, the proposed hybrid approach is expected to outperform rule-based, supervised-only, offline-RL-only, and prompt-only agent baselines in composite decision quality and robustness while maintaining substantially better latency and cost than cloud-only frontier-model pipelines.
