Not every clinical agent needs an LLM.
Jev is a System One model. Instead of writing prose, it answers structured questions, yes/no and multiple choice, with calibrated probabilities. Deterministic rules turn those answers into findings. Fast, cheap, and easy to audit.
ICD-10 Code Reviewer · 40 inpatient stays · synthetic data · no PHI
inpatient stays reviewed
Jev requests
total cost of the run
p50 per request
A different kind of model for a different kind of work.
Inpatient coding review means hundreds of small judgments per stay. Is the condition current? Is the stage documented? Is there MEAT, that is, was it monitored, evaluated, assessed or treated? LLMs can make those calls, but at that volume they are slow, costly and hard to audit.
Jev answers each of those questions as a structured prediction with a probability attached. Coding rules, written as code and reviewable as code, turn the answers into findings. There is no LLM anywhere in the pipeline, and Jev never sees an ICD-10 code.
The ICD-10 Code Reviewer.
The first agent on Jev checks a coder's draft codes against the whole chart, with evidence cited for every finding. Nothing is auto-assigned. The coder stays in charge.
- Pull the stay
The Kartha MCP FHIR server provides the encounter, the draft codes, labs and every clinical note including the discharge summary. Identifiers are removed from note text before anything is sent to Jev.
- Ask Jev small questions
Hundreds per request: is it current, what stage or side, what caused it, is there MEAT, was it present on admission, is anything treated but not coded.
- Rules build the report
Coding rules turn Jev's answers into findings. Each one cites the exact sentence it came from.
Unsupported codes
Draft codes with no supporting documentation in the chart.
History coded as active
Resolved or historical conditions carried forward as current.
Missing specificity
N18.9 where the notes support N18.32. Combination codes such as E11.22.
Documented but uncoded
Conditions in the notes with no code, flagged with CC/MCC and HCC impact.
POA and sequencing queries
Present-on-admission and principal-diagnosis questions for the coder to resolve.
Evidence on every line
The sentence, the note, the timestamp. Reviewable by a coder in seconds.
One platform, both kinds of agent.
LLM-powered agents
Chat and reasoning agents built with the Agent SDK, fifty clinical skills, and full receipts on every answer. Best when the question is open and the clinician is in the loop.
Purpose-built models
Agentic workflows on models like Jev, at millisecond speed and a fraction of the cost. Best when the work is hundreds of small structured judgments. Same MCP FHIR server, same guardrails.
A word on PHI. The reviewer removes identifiers from note text before anything is sent to Jev, but redaction alone is not a compliance plan. For real patient data, plan for a BAA with the model provider, or run an open-source variant of Jev in your own cloud.