DenialDefender

US hospitals leave hundreds of billions in recoverable revenue on the table every year because writing insurance appeal letters is too slow and expensive to scale. DenialDefender is the autonomous appeals layer designed for hospital revenue cycle management (RCM) teams and third-party RCM services. A user uploads a denial letter, and our multimodal system ingests the document, the patient's full chart, the payer's medical policy, clinical literature, and past successful appeals. Within 60 to 90 seconds, it drafts a complete, submission-ready appeal packet, including a medical necessity letter, code corrections, citations, and an overturn confidence score. Because this is a massive memory-bandwidth-bound workload requiring context windows up to 64K tokens for the demo (scaling to 200K+ in production), we rely on the AMD Instinct MI300X. We run both a primary reasoning model (Qwen3-32B) and a vision model (Qwen2.5-VL-7B) co-resident on a single GPU in FP16. This drives our end-to-end compute cost down to just $0.03 per appeal, enabling our core contingency business model: we don't charge if we don't recover.