01
Extract
Turn chart text into cited facts
A language model structures the chart into atomic clinical facts, each linked to the evidence that supports it.
It cannot select a code, modifier, or unit.
Built for orthopedic professional coding
Every code traces back to a chart.
OrthoCode AI extracts clinical facts, makes bounded coding decisions, and enforces CMS requirements deterministically—so every proposed line has a defensible trail.
Open demo
In development · human review is mandatory
Licensed
Versioned codebooks
Bounded
Typed decisions
Deterministic
CMS rule enforcement
The problem
Orthopedic coding leaves very little room for a plausible guess.
Multi-procedure claims, NCCI bundling edits, and modifier disputes make orthopedic surgical billing unusually exposed to denials and audits.
NCCI
Bundling edits
−59 / X{EPSU}
Modifier disputes
MUE
Unit limits
This is a domain where hard rules and complete evidence trails matter more than raw automation speed. The architecture should reflect that.
How it works
Three layers. Each with a strict boundary.
JEV—TypeSafe's bounded decision engine—sits between fact extraction and deterministic rule enforcement. That separation is the safety model.
02
Decide
Choose from bounded candidates
TypeSafe JEV weighs only candidates retrieved from the active, licensed and versioned codebook release.
It cannot invent a code outside the licensed release.
03
Enforce
Apply rules that cannot be overruled
Deterministic Python checks enforce active-code, NCCI, MUE, add-on and fee-schedule constraints.
No model can override a hard CMS rule.
Trust by design
The system is designed to make uncertainty visible.
No inflated autonomy claim. No mystery score standing in for evidence. No model opinion overriding a hard rule.
Read the security postureFails closed
If the decision engine is unavailable or unsure, the chart is routed to a human. No silent fallback, no fabricated lines.
Evidence stays attached
Every proposed line retains its cited chart evidence and decision trace for reviewer and audit use.
Humans remain in control
Autonomous coding is disabled by default. Corrections are append-only and feed controlled evaluation.
Bounded data sharing
Development is de-identified only; JEV receives bounded evidence spans and PHI sharing is off by default.
Questions, answered plainly
What a careful coding team should ask.
Is this autonomous coding?
No, not by default. Autonomous submission is disabled unless explicitly enabled and every measured threshold is met; a human reviews every chart today.
What happens if the AI is unsure?
The chart fails closed to human review. The system preserves the evidence and candidates it has, but it does not guess to avoid an empty result.
Where do the codes come from?
Only from the active, licensed and versioned CPT, HCPCS and ICD-10 codebook release for the service date—never from free-form generation.
What is JEV?
TypeSafe JEV is a bounded decision engine. It chooses among retrieved candidates and can abstain; it never generates a new billing code. Its probabilities are routing signals, not substitutes for coder judgment.
Built carefully, not loudly