SwasthVision- Machine Learning-Based Indian Food Calorie Estimator with Mood-Linked Notifications

Abstract

SwasthVision is a novel mobile health (mHealth) system designed to resolve the inherent conflict between advanced AI intelligence and individual data sovereignty. Addressing the "privacy crisis" in conventional cloud-first nutritional applications, the system implements a Local-First Dual-Tier Hybrid AI Engine.The architecture intelligently partitions computational tasks into two layers: Tier 1 utilizes a lightweight 135M-parameter Small Language Model (SmolLM-135M) running entirely on-device for immediate, offline health queries; Tier 2 offloads complex nutritional reasoning to a local high-performance hardware bridge (a home PC/Laptop) running Ollama and Meta’s Llama 3.2-3B model. By establishing a secure WiFi/LAN bridge to utilize NVIDIA CUDA acceleration (896 cores), the system reduces AI response latency from 5–15 seconds on a mobile CPU to just $76\text{ ms}$, achieving cloud-level performance without data egress to the public internet. Core functional features include a "skip-login" accessibility model to prevent personal identification and a local-first persistence layer using SQLite with FTS5 Full-Text Search for instantaneous food mapping [1, 1]. All health recommendations are grounded in Explainable AI (XAI) logic, specifically utilizing clinical formulas like the Mifflin-St Jeor equation to ensure every dietary target is traceable and auditable. Comparative validation demonstrates that SwasthVision offers 100% offline functionality and zero data privacy risk, providing a robust, risk-aware framework for the next generation of responsible health assistants.