Australian AI Care Partner Heidi now automates clinical documentation across more than 190 countries, supporting roughly 2.7 million patient interactions per week. The scale owes less to model improvements than to infrastructure decisions made years before global expansion, according to co-founder and chief technology officer Yu Liu.
Healthcare carries different engineering demands than other markets. "In most industries, an AI feature that is wrong two percent of the time registers as an inconvenience, while in healthcare that same error rate becomes a clinical safety issue," Liu said. "The architecture has to be built around the assumption that every output may be scrutinised, audited, and relied upon in a patient's care."
Compliance is architecture, not a promise
Heidi runs fully logically isolated production deployments across the world, so data stays in-region by design. Clinicians in Sydney, London, Tokyo, or Denver operate under different rules - the Australian Privacy Principles, GDPR, APPI, and HIPAA - and each jurisdiction requires patient data to remain within its borders.
Auditability is built in from day one. The organization must be able to answer, months later, what the model saw, what it produced, and what the clinician changed for any given session.
"The blast radius of change must be engineered down," Liu said. "In less regulated industries you can ship fast and fix forward, but in healthcare we invest heavily in making change safe by default, with continuous integration gates on risky change classes, canary releases, and treating even database schema and index changes as code that goes through review."
Why a document database fits clinical AI
Heidi consolidates medical data from forms, referrals, and clinician notes into one format and location. An AI Scribe session isn't a single blob - it's transcripts, structured notes, templates, documents, patient context, and EHR integration state, all of which change week to week. Rigid rows and columns suited that workload poorly.
MongoDB Atlas gives the team a document model that keeps session data together and supports AI workflows without a migration freeze. Liu said migrating to Atlas reduced latency on key APIs by nearly 33%. The company also uses Atlas Vector Search so it doesn't need a separate vector database.
"The model is maybe 20% of the system, and the data architecture is what determines whether the other 80% holds up under real clinical load," Liu said.
Jurisdiction-aware treatment for a RAG system
Heidi's Evidence feature retrieves from licensed clinical knowledge bases - BMJ Best Practice, NICE CKS, and MIMS - and is jurisdiction-aware. A U.K. clinician gets U.K. guidance; an Australian clinician gets Australian formularies, because the right answer in one country can be the wrong answer in another.
"Retrieval is a data architecture problem before it is an AI problem," Liu said. "In consumer RAG, you retrieve from the open web and hope, whereas in healthcare what you retrieve from is the compliance surface."
Embeddings and vector indexes live in the same regionally isolated deployments as the rest of the data. Retrieval physically cannot cross a residency boundary, and citations are a hard contract rather than a prompt suggestion.
Professionals looking to gain relevant skills in these systems may find value in AI for Healthcare training resources.
Multi-region scale with a lean team
Expansion into the U.S. market came without re-engineering. "Entering the U.S. market meant standing up another region on rails we had already built rather than re-engineering for HIPAA after the fact," Liu said.
Beth Israel Lahey Health, one of New England's largest health systems, rolled out Heidi's AI scribe following a pilot where 74% of clinicians reported reduced after-hours documentation. Rural system MaineGeneral Health also selected Heidi as a strategic partner.
Liu's advice for anyone building a data-heavy AI product: plan for growth early. "Re-partitioning a large, hot, always-on collection is a serious engineering program, whereas choosing a shard key on day one is a design meeting."
Why this matters for healthcare professionals
AI tools in clinical settings are only as good as the infrastructure under them. Understanding the difference between safe-by-default design and processes that ship fast, fix forward may shape which products clinicians choose - and avoid.
A platform that works in a patient's literature tells you little if it can't hold residency, retrieval, or audit requirements. And since trust is built on reliability, not just performance, professional in healthcare should consider a vendor's architecture as seriously as they weigh its marketing. Get used to asking where your data lives, how changes are tested, and what happens if the model is wrong, because those answers shape patient safety.
"Clinician trust is the product, and trust is architectural," Liu said.
Your membership also unlocks:
