Algorithm
Hybrid Contextual Bandit
Reinforcement learning with two-level exploration-exploitation. Your profile informs program
selection (Level 1); program response informs exercise substitution (Level 2).
Shared parameters let exercise feedback improve program recommendations and
vice versa.
Based on Li et al. 2010
Safety
ACWR-Guarded Progression
Every workout passes through an Acute:Chronic Workload Ratio guardrail that
matches sports-science research on injury prevention. We slow progression when
your training load exceeds recovery capacity — no pushing through red flags.
Based on Gabbett 2016
Matching
22-Dim User Embedding
Your profile becomes a 22-dimensional vector covering goals, experience,
equipment, schedule, injuries, and history. We match against a library of
program embeddings using cosine similarity in pgvector for sub-millisecond
retrieval at scale.
Pure vector search — no keyword matching
Coaching
RAG-Grounded Companion Chat
When you ask about muscle soreness, the coach retrieves relevant
exercise science from a curated knowledge base via hybrid BM25 + semantic
search, then generates a response grounded in that context. No hallucinated
medical claims.
Retrieval Augmented Generation
