Med Audit Swarm

Med-Audit-Swarm is an autonomous, multi-agent AI safety layer designed to intercept fatal prescription errors and eliminate clinical "alert fatigue." In modern hospitals, legacy Electronic Medical Record (EMR) systems overwhelm physicians with minor, rule-based warnings, causing doctors to instinctively ignore alerts—even the life-threatening ones. Furthermore, relying on standard Generative AI to solve this creates a massive liability, as single-model LLMs frequently "hallucinate" or guess clinical data to fulfill a prompt. To solve this, I built a deterministic digital assembly line using the CrewAI framework. When a physician uploads an unstructured patient EMR PDF, the system deploys a specialized three-agent swarm: The Clinical Pharmacist: Ingests the messy text and extracts structured data (current medications, allergies, disease history) with 100% precision. The Toxicology Auditor: Cross-references this extracted data against actual pharmacological science to flag lethal cross-reactivities and contraindications. The Senior Reviewer: Acts as a strict quality-control gate, scrubbing the findings for false positives to ensure the physician only sees critical, highly accurate warnings. Under the hood, this system relies on strict "Zero-Hallucination" engineering, locking model temperatures near zero to guarantee fact-based, explicit outputs without AI guessing. Crucially, clinical AI cannot compromise patient privacy. Med-Audit-Swarm is architected specifically for the AMD Instinct MI300X pipeline. By utilizing its massive 192GB HBM3 memory alongside the vLLM engine, I am able to run the massive Qwen 2.5 72-Billion parameter model entirely on-premise. This ensures strict HIPAA compliance and data sovereignty while reducing a 15-minute manual chart review into an autonomous, 5-second audit.