Healthcare data workflows

Convert HL7 v2 messages to CSV, for analytics and ops.

A file of ORU results or ADT events is full of answers your team needs in a table, not in pipe-delimited segments. Describe the rows: one per message, or one per OBX observation, with the fields that matter. Data Shepherd writes reviewable scripts that parse the segments into typed columns, and every run after approval is sealed, with nothing sent to a model.

Free plan available. No card required.

No data sent to a model at run-time · transparent, versioned scripts
HL7 v2 to CSV
Input · ORU^R01 (synthetic)
MSH|^~\&|LAB|HOSP|EHR|CLINIC|202606051200||ORU^R01|0001|P|2.5
PID|1||MRN000123^^^HOSP||DOE^JANE||19800101|F
OBX|1|NM|GLU^Glucose||96|mg/dL|70-99|N
OBX|2|NM|HBA1C^Hemoglobin A1c||5.4|%|4.0-5.6|N
▾ AI-written transformation, reviewed and versioned
Output · results.csv
patient_id,test,value,unit,ref_range,flag
MRN000123,Glucose,96,mg/dL,70-99,N
MRN000123,Hemoglobin A1c,5.4,%,4.0-5.6,N
Runs with no AI, your data stays out of the model

How it works

From a messy file to a tested pipeline, in minutes.

1

Describe it and bring your spec

Plain English, plus a sample, a spec, or a data dictionary. Cannot share data at all? Build from the spec alone.

2

AI writes transparent scripts

You get reviewable, versioned Python, never a black box. Inspect every line and preview the output on a sample.

3

Run it anywhere, with no AI

Trigger by API, schedule, or click, or pull from SFTP, S3, or Azure Blob. The sealed script does the work; nothing goes to a model.

The AI builds it. A sealed script runs it.

Most AI data tools push records through a language model on every run. We keep the model in the build phase only, so production never sends your data to a model.

Build (one time): AI writes the scripts, from your spec, or a sample you control
Every run after: sealed sandbox, no network, no AI, no data to any model

Full data-flow table and sub-processor list in our Trust Center →

Common questions

HL7 v2 to CSV, answered.

Our file mixes message types. Is that a problem?

No. Describe how each type should be handled, parsed into separate outputs, filtered out, or flagged, and the generated script does exactly that, the same way on every run.

How are repeating OBX segments handled?

Your choice: one row per observation with patient fields repeated is the common shape for analytics, but you can describe any structure and preview it on a sample before running the full file.

Is PHI sent to an AI model when this runs?

No. Runs execute a sealed, integrity-checked script in a sandbox with no AI and no network. The model is only involved while the scripts are written, and you control what it sees, down to sharing only a spec.

Working with claims or clinical data? See the dedicated healthcare page →

Describe it once. Run it forever.

Build the transformation on a sample in minutes, review the scripts line by line, and let the sealed runs handle every file after that.