Healthcare software built on HL7 FHIR under HIPAA and GDPR
Most healthcare work sits between systems that were never designed to talk to each other. An EHR integration usually means HL7 v2 ADT and ORU feeds from the hospital's interface engine on one side and FHIR resources on the other, so we build the translation once and expose one FHIR API for portals, telehealth and analytics. Patient identity is resolved through a master patient index rather than name matching, because duplicate records are where clinical errors start. Access control maps to clinical roles, with break-glass access for emergencies that is logged for review. Consent, retention and residency are stored as data attached to each record, which is what lets a GDPR subject access request or a HIPAA audit be answered from the system rather than reconstructed by hand. Pharma CRM and ERP work follows the same rule.
Clinical AI kept defensible with local validation and human oversight
The AI that pays off fastest in healthcare and pharma is rarely the headline diagnosis. It is the queue: prioritising an imaging worklist so the scan that needs a radiologist first is read first, reading referral letters and claims into structured fields, or flagging a patient whose observations have been drifting since the last ward round. We keep that defensible by validating every model on a held-out set drawn from your own population and equipment, because a model that scored well on a public benchmark often degrades on a different scanner or a different case mix. Thresholds are set with your clinicians, who decide how to trade missed cases against false alarms. Each prediction is logged with its model version and inputs, so a flagged or unflagged case can be reconstructed months later for a clinical governance review.
Interoperable patient data in production with audits that stay routine
Going live in a hospital or a pharma operation is never a single switchover. We run the new system alongside the old one first, with interfaces in shadow mode and reconciliation reports that compare records between the two until the numbers match, then move clinicians across department by department. Hosting follows the data: a region-pinned cloud tenancy where residency rules allow it, or on-premise inside the hospital network where they do not, with backups restored on a schedule to prove they work. The failure modes we monitor are unglamorous, such as a stalled HL7 queue or an interface silently dropping messages, because those are what create the duplicate records the project set out to remove. You are left with documented interfaces, a data dictionary, runbooks and audit evidence that comes out of the system itself.