Powered by MongoDB

How Heidi Built a Globally Scalable, Compliance-Ready AI Platform for Healthcare

For organizations operating in highly regulated industries such as healthcare, financial services, and transportation, building accurate, secure, and reliable artificial intelligence is a major engineering challenge. Developing AI-powered products is even more complex in sectors where regulations require organizations to adopt new technologies carefully and methodically. As a result, many companies are modernizing their data infrastructure to meet the growing demands of AI.

Australian-founded AI care partner Heidi is an example of successful healthcare technology modernization. Its flagship product, Heidi Scribe, automates many time-consuming administrative tasks for clinicians in more than 190 countries and supports approximately 2.7 million patient interactions each week. According to Yu Liu, Heidi’s co-founder and chief technology officer, the company’s global expansion is rooted in infrastructure decisions made long before it reached its current scale.

“In most industries, it would be inconvenient if an AI function were wrong 2% of the time, but in healthcare, the same error rate becomes a clinical safety issue,” Liu says. “The architecture must be built on the premise that every output can be scrutinized, audited, and trusted in patient care.”

Why Production AI in Healthcare Requires a Different Architecture

For Heidi, data residency is not an optional feature—it is a prerequisite. Clinicians in Sydney, London, Tokyo, and Denver operate under different regulatory frameworks, including the Australian Privacy Principles, GDPR, APPI, and HIPAA. Patient data must remain within the appropriate geographic region to support compliance and protect sensitive health information.

Heidi operates logically isolated production environments around the world, ensuring that data remains within its designated region. Auditability is also built into the platform from the beginning. Healthcare organizations need to be able to determine, months later, what data an AI model accessed, what it produced, and what the clinician ultimately changed.

“We need to reduce the blast radius of change,” Liu says. “In less regulated industries, teams can ship quickly and fix issues later. In healthcare, we invest heavily in making changes safe by default. That includes risk-based change classifications, continuous integration gates for canary releases, and treating database schema and index changes as code that must pass review. Our speed is a product of that safety, not something achieved in spite of it.”

Choosing a Database for AI-Powered Healthcare Workflows

Heidi processes diverse medical data from multiple sources, including forms, referrals, clinician notes, transcripts, templates, and electronic health record integrations. The company needed to bring this information together in a consistent, accessible environment that could connect seamlessly with AI workflows. Traditional databases built around fixed rows and columns are not always well suited to this type of dynamic workload.

These requirements made a document database a natural choice for Heidi. MongoDB provides the flexibility needed to support rapidly changing AI data models without requiring the underlying database to be rebuilt each time the product evolves.

“The model is probably 20% of the system, and the other 80% is the data architecture that determines whether it can withstand real-world clinical loads,” Liu says.

An AI scribe session is not a single data point. It is a collection of transcripts, structured notes, templates, documents, patient context, EHR integration statuses, and other related artifacts that change over time. MongoDB enables Heidi to store this session data in a structure that reflects how clinicians work while allowing the company to evolve its data model without pausing development for extensive migrations.

MongoDB Atlas stands out because it combines the scalability, flexibility, and performance of a document model with AI-enabled capabilities such as MongoDB Vector Search. “This means Heidi does not need to add a separate vector database to extend its existing platform,” Liu says.

With availability across more than 130 cloud regions, as well as on-premises and hybrid deployment options, MongoDB Atlas is a globally distributed database platform designed for organizations with complex data residency requirements. Its unified query API allows developers to build full-text search, real-time analytics, and event-driven applications without adding unnecessary architectural complexity.

“Heidi Scribe transforms large volumes of medical documents into vector embeddings through LangChain in Atlas, enabling semantic search that directly connects transcribed medical terms with relevant external knowledge,” Liu adds. “When we moved to Atlas, we reduced latency on our core APIs by nearly 33%.”

Building a Reliable Clinical RAG System

“Search is a data architecture problem before it is an AI problem,” Liu says. “In consumer RAG, you may search the open web and hope the results are relevant. In healthcare, you search within the compliance surface.”

Heidi Evidence draws from accredited clinical knowledge bases, including resources from partners such as BMJ Best Practice, NICE CKS, and MIMS. The platform is also jurisdiction-aware, so clinicians in the United Kingdom receive UK-specific guidance while clinicians in Australia receive information relevant to Australian practice. This distinction is essential because guidance that is correct in one country may be inappropriate in another.

Heidi’s embeddings and vector indexes reside in MongoDB Vector Search within the same geographically isolated deployment as the rest of the organization’s data. This approach helps prevent information from being physically retrieved outside its designated residency boundaries and avoids the need to operate a separate vector database with its own security and compliance requirements.

Because the AI model references retrieved content that is already linked to source records, citations function as verifiable connections to underlying information rather than unsupported suggestions. This helps improve transparency and trust in clinical AI workflows.

Regional Isolation Supports Global Healthcare Compliance

“Each region is a completely isolated production environment, with its own MongoDB Atlas cluster, compute resources, and encryption keys,” Liu says.

“That allows us to enter a US health system, an NHS trust, or a hospital group in Australia and provide a clear answer about data residency. Residency rights are not simply promised contractually—they are enforced by the infrastructure,” he explains. “Running multiple isolated regions with lean teams works because the database layer is managed and consistent. We are also multicloud, which means a new region can be deployed on top of the rails we have already built.”

This architecture is particularly evident in the United States, where Beth Israel Lahey Health, one of New England’s largest health systems, implemented Heidi’s AI scribe following a pilot. The assessment found that 74% of clinicians reported spending less time on after-hours documentation, often referred to as “pajama time.” Nonprofit health system MaineGeneral Health also selected Heidi as a strategic partner for rural healthcare operations.

“Entering the US market meant launching another region on top of the infrastructure we had already built, rather than redesigning the platform for HIPAA compliance after the fact,” Liu says.

Lessons Learned and the Future of Healthcare AI

“Repartitioning a large, high-volume, always-on collection is a full-fledged engineering program, but choosing shard keys on day one is a design meeting,” Liu says. “We are currently partnering with MongoDB to do that work, but the lesson for anyone building data-intensive AI products is that horizontal scalability for rapidly growing data is as much a founding decision as data residency.”

Heidi is expanding beyond consultation notes to support the broader clinical workflow, including pre-consultation preparation, post-consultation documentation, referrals, and workflow automation. The company is also exploring how MongoDB, large language models, and proprietary tools can support an agent ecosystem for clinical workflows.

“In healthcare AI, reliability engineering is trust engineering,” Liu says. “Clinician trust can be lost as quickly as a bad memo through downtime, latency, or data inconsistency. Some of the most valuable work is invisible, including canary releases with automatic rollbacks, continuous integration gates for database changes, and cross-region consistency checks. Clinician trust is the product, and trust is the architecture.”

Sponsored articles are content created by companies that pay us to post or have a business relationship with VentureBeat, and are always clearly marked. For more information, please contact [email protected].