Software engineering for healthcare
Patient data protected by regulation. Systems that cannot go down. Networks that are closed by policy. Legacy platforms that every new system has to work with.
Interoperability
Every new system has to talk to systems that were there first, and those systems set the terms. In practice that means HL7 v2 from older systems, FHIR R4 on newer ones, CCDA where clinical summaries are exchanged, and Epic interfaces where the health system runs Epic. Getting the message in is only the first part. It has to resolve to the right patient, land on the right record, and trigger the next step.
Compliant cloud
HIPAA affects the architecture from the beginning. Encryption in transit and at rest, including backups and logs. Least privilege for people and services. Segmentation between the public tier and the database. Tamper evident audit trails, tested backups, and only the services covered by the business associate agreement touching protected health information.
Running where the data is
Some networks have no internet route because of policy. A cloud native platform can still run there, but services that normally come from the outside have to be handled locally. Identity, storage, telemetry, update checks, licence checks and other outbound calls all have to be accounted for before the system can run in that environment.
AI in the clinical workflow
Software can measure consistently and process more than a person can review by hand, but the clinician still makes the decision. The review step belongs in the product. That can be a flag, a queue or a threshold that sends the result to a person, with the model version recorded alongside it.
Clinical systems have to keep running
That changes the migration from the start. The new environment runs alongside the old one, data stays synchronised, integrations are tested while the old system is still serving users, and a rollback path is prepared before the switch. With multiple facilities, each site can have its own schedule and its own legacy systems.
Identity is another constant. Data comes from several systems and several formats, and it has to resolve to the right person before it reaches a table. On an oncology EHR data pipeline, a master patient index was built before the parsers. The extract and load code covered ten message types across HL7 v2, two CCDA variants and flat report exports, and every raw export was retained in object storage so a row could be traced back to the file it came from.
We built a HIPAA compliant AWS environment for a multi site hospital network and moved them onto it without interrupting clinical operations. Each facility moved on its own schedule, with real time FHIR R4 exchange across the sites.
Scheduled exports, a landing layer that keeps the raw file, parsing across HL7 v2 and CCDA, an identity layer, and a database the data team can query.
On a patient platform for allergic contact dermatitis, web, iPhone and Android, with an ingredient reference table cleaned by language models and every new mapping held for review before it reached a patient's safe list.
Slide viewers, scoring, annotation and the automation and QA around them, with the specialist reviewing the result and the software handling the measurement.
A health system needs to move to a compliant environment without interrupting care. A data organisation has exports it cannot query. A platform needs to reach the record clinicians already work in rather than sit beside it. Or the system has to run inside a network that has no route out.
Questions people ask
Can you migrate a live hospital network without downtime?
That is how we handle it. The migration runs alongside the existing system, data stays synchronised, integrations are tested against the new environment, and a rollback path is ready. Facilities can move on their own schedules.
Do you work with HL7 v2 or FHIR?
Both. Most real deployments need both. FHIR R4 on the modern side, HL7 v2 from legacy systems, and CCDA where clinical summaries are exchanged.
What happens when an AI result is wrong?
There is a review step before the result is used. The review can be a flag, a queue or a threshold that sends the result to a person. The model version is recorded with the result.
Moving clinical systems or data without interrupting care?
Tell us what the systems are, where the data sits, and who depends on it.
Start a conversation