Rebound
From Setback to Comeback with AI
The problem Rebound solves
Athletes recovering from injuries often face uncertainty regarding diagnosis, treatment options, and return-to-play timelines. Traditional rehabilitation methods rely on generalized recovery plans that may not account for individual conditions, history, or affordability. Rebound addresses these challenges by providing an AI-powered system for real-time injury detection, personalized rehabilitation guidance, and treatment cost optimization.
It enables faster and more accurate diagnosis by using Transfer learning: CNN-based models to analyze MRI and X-ray images while integrating past injury records and patient-reported symptoms for a comprehensive assessment.
The RAG-powered fitness coach chatbot offers personalized recovery guidance by continuously ingesting and indexing new research, patient cases, and medical literature to provide up-to-date recommendations.
Rebound also optimizes treatment plans through a reinforcement learning model that dynamically adjusts costs based on affordability, success rates, and athlete-specific factors, ensuring cost-effective and high-success-rate treatment plans.
Additionally, the system includes an XGBoost-based model that estimates return-to-play timelines by analyzing injury severity, treatment response, and past recovery data.
By combining AI-driven diagnosis, rehabilitation, and cost optimization, Rebound helps athletes make informed decisions, recover efficiently, and return to play with confidence.
Challenges I ran into
One of the biggest challenges was building the RAG-powered fitness coach chatbot to provide accurate and timely rehabilitation guidance. Since the chatbot continuously ingests new research and patient data, ensuring reliable, context-aware responses was difficult. One major issue was hallucination, where the model generated misleading or incorrect medical advice.
Another key challenge was latency and scalability. Handling large-scale medical literature updates while ensuring low-latency responses required optimizing the indexing pipeline.
Another challenge was ensuring accurate injury detection with CNNs. Variability in medical imaging and dataset biases led to inconsistent model performance. To improve accuracy, we applied data augmentation, transfer learning with pre-trained medical models, and domain-specific fine-tuning using high-quality datasets.