Engineering
How Deosai actually works.
For the technical evaluators on your team. The architecture, the data, and the testing discipline behind autonomous coding and clean claims.
By the numbers
0+
encounters processed
0+
payer-specific rules
15GB
clinical corpus
0
encounter regression suite
Architecture
A system you can put under diligence.
Hybrid AI + rules
We pair large language models with an explicit, versioned rules engine. The model proposes; the rules constrain. Payer edits, NCCI logic, and modifier policy live as code we can read, test, and audit — not as opaque weights.
Grounded in production data
Recommendations are grounded in a 15GB clinical corpus and 90,000+ real ambulatory encounters, with 23,421 reconciled claim relationships and 10,300 adjudicated claim lines as ground truth.
Regression testing on every release
A 4,000-encounter regression suite runs before any model or rule change ships. We measure coding accuracy and claim outcomes against known-good labels, so a release can only go out if it holds or improves the bar.
Explainability
Every code carries a citation back to the documentation that supports it, plus the rule that allowed it. A human coder can see exactly why a recommendation was made before approving it.
Human-in-the-loop by design
Deosai never bills unattended. Coders review and approve, and every action is captured in an audit trail with role-based access control.