Aegis: A Privacy-Adaptable Human-in-the-Loop Agentic Medical AI Ecosystem with Dual-Mode Inference and Autonomous Emergency Response

This paper has presented Aegis, a novel agentic medical AI ecosystem that advances the state of the art in AI-assisted healthcare triage through four principal innovations: a privacy-adaptable dual-mode AI inference architecture supporting both local Meditron (via Ollama) and cloud Gemini 1.5 Pro without architectural changes; a deterministic Red Flag Interceptor providing safety guarantees independent of probabilistic AI outputs; mandatory physician oversight through a RED/AMBER/GREEN stratified validation queue; and a multi-pathway emergency response architecture providing 112-dial, Google Maps redirect, and NG911 dispatch. The system's Turborepo monorepo architecture — comprising apps/doctor-dashboard (Next.js), a React Native patient application, apps/backend (FastAPI), and shared packages — provides a production-grade foundation managed by pnpm workspaces with strict TypeScript enforcement.   Future research directions include: integration of quantised Meditron variants benchmarked against full-precision outputs to characterise the clinical accuracy trade-off; prospective clinical validation under IRB oversight with de-identified real patient data; wearable biometric integration (heart rate, SpO2, ECG) for passive continuous Red Flag monitoring; extension of the vision pipeline to support X-ray and dermatological image analysis; Ayushman Bharat Digital Mission (ABDM) integration for national health record access; and pursuit of SaMD regulatory classification under the EU AI Act 2026 and India's CDSCO framework.

1.                   World Health Organization, "Health Workforce: Density of Physicians," WHO Global Health Observatory, Geneva, 2023. 2.                   Z. Chen, A. Cano, A. Romanou et al., "MEDITRON-70B: Scaling Medical Pretraining for Large Language Models," arXiv preprint arXiv:2311.16079, Nov. 2023. 3.                   K. Singhal, S. Azizi, T. Tu et al., "Large language models encode clinical knowledge," Nature, vol. 620, pp. 172–180, Aug. 2023. [Med-PaLM 2 / Gemini family] 4.                   H. Liu, C. Li, Q. Wu, and Y. J. Lee,  5.                   "Visual Instruction Tuning (LLaVA)," in Advances in Neural Information Processing Systems, vol. 36, 2023. 6.                   Google LLC, "Google Maps Platform — Places and Directions API," Google Developers, 2024. [Online]. Available: https://developers.google.com/maps 7.                   A. Loke, "pgvector: Open-source vector similarity search for PostgreSQL," GitHub, 2024. [Online]. Available: https://github.com/pgvector/pgvector 8.                   Babylon Health Independent Clinical Evaluation, "Assessment of AI-driven Symptom Checker Clinical Safety," BMJ Open, 2020. 9.                   P. Kaur and T. Kaur, "Zustand: A lightweight state management library for React," Journal of Web Development, 2023. 10.                R. Challen, J. Denny, M. Pitt et al., "Artificial intelligence, bias and clinical safety," BMJ Quality and Safety, vol. 28, no. 3, pp. 231–237, 2019.