Zabber · AI fitness coach
A flat exercise list, turned into an injury-aware coach's brain
Names, target muscles, equipment: not enough to coach anyone safely. We built the enrichment pipeline and the retrieval layer the coach runs on.
1,324
Exercises enriched and embedded
339
Carry an injury-risk flag
718
Identified as home-friendly
15
Structured fields per exercise, plus an embedding
The pipeline
- 01
Rules first
Equipment, home-friendliness and one-sided moves, derived in code. No model sees a fact.
- 02
AI classification
Against a fixed schema: difficulty, movement, force, impact, injury risk, goals, rep ranges.
- 03
Name-free embeddings
Embedded from structure, not name, so similar-sounding moves stop matching.
- 04
Versioned, resumable runs
Runs resume where they stopped. New exercises are enriched on demand.
Inside the enriched dataset
Difficulty
Movement type
Injury-risk flags by joint
How the data powers the coach
Injury-aware search
Exercises loading an injured joint are excluded in the database query.
Plan validation
An unsafe or too-advanced pick is swapped for the nearest safe one.
One-step swaps
Easier, harder, no-equipment or joint-friendly, computed with no AI call.
"For you" picks
From the embedding of what each person actually trained.