Problem Statement: The "Silver Tsunami" & The Trust Gap By 2030, one in six people globally will be over 60, facing a trifecta of threats: Physical Danger (falls), Cognitive Decline (memory loss), and Financial Exploitation (medical billing fraud). ● The Failure of Current AI: Standard LLMs often hallucinate medical advice or fail to detect subtle billing errors. ● The Opportunity: We need an agent that doesn't just "chat," but sees hazards in real-time and reasons with the rigor of a clinician before authorizing actions.
- System Architecture: The Tri-Agent Mesh Our solution orchestrates three specialized agents using Google Antigravity: Agent A: MemPal Live (The Senses) ● Role: Real-time environmental monitoring and memory augmentation. ● Tech Stack: Gemini Live API (WebSockets) + Browser Text-to-Speech (TTS). ● Workflow: Streams video/audio at 3FPS. Uses Gemini’s multimodal vision to detect hazards (e.g., "Wet floor") and immediately interrupts the user via TTS ("Stop! Hazard ahead."). It also serves as an external memory bank for dementia patients, recalling context from visual history. Agent B: Medi-Guard (The Brain) ● Role: Clinical validation of symptoms, prescriptions, and invoices. ● Tech Stack: Gemini 3 Pro (thinking_level="high") + OpenMed RAG. ● Innovation: We ground the model in a curated slice of the OpenMed (Medical-Reasoning-SFTMega) dataset. When a user uploads a medical bill, the agent cross-references the treatment (CPT codes) against the diagnosis (ICD-10) using OpenMed’s chain-of-thought logic. If a clinical mismatch is found (e.g., "MRI for a sore throat"), it flags the bill as potential fraud. Agent C: Peer Aid (The Hands) ● Role: Autonomous execution of verified tasks. ● Tech Stack: Tool Use (Function Calling) + Thought Signatures. ● Workflow: Once Medi-Guard verifies a claim with >95% confidence, Peer Aid utilizes tool calling to interface with simulated banking APIs, releasing funds directly to providers without user administrative burden.
- Technical Implementation & Novelty ● Grounding Strategy: To bypass context window limits while maintaining accuracy, we use Retrieval Augmented Generation (RAG) on a high-density "Expert Slice" of the OpenMed dataset, converting 2.22B tokens of raw data into a targeted knowledge base for the agent. ● Deep Reasoning: We explicitly enable Gemini 3’s Deep Think capabilities for the Medi-Guard agent, ensuring it explores multiple hypotheses for medical validity before rendering a verdict. ● No-Code Orchestration: The entire frontend and logic flow were architected using Google AI Studio’s "Build" Mode (Vibe Coding) and managed via Google Antigravity, demonstrating the power of agentic coding to accelerate complex healthcare software development.
- Impact Potential ● Safety: Reduces fall risk through proactive, sub-second visual warnings. ● Economic: Prevents thousands of dollars in medical fraud per user by validating every line item against clinical standards. ● Scalability: The software-only architecture (requiring only a smartphone) makes high-quality care accessible to the 13 million+ elderly population in developing regions like Pakistan.