This article originally appeared as a paid post on the South China Morning Post, sponsored by Amazon Web Services (AWS) Hong Kong.
A hospital Radiology Department in Hong Kong manages a continuous stream of patient inquiries at all hours, across more than 1,000 distinct examination items with pricing logic that varies by modality, contrast, anatomy, and applicable package. Referral letters arrive as photos containing handwritten notes and mixed Chinese-English text. Manual processing alone could not keep up, and hiring was not the fix the department could reach for indefinitely. The department selected Vascue, deployed on AWS infrastructure in Hong Kong, to take over the routine intake and quotation work, and set three requirements from the outset: compliance with hospital standards, proper data handling, and no disruption to the clinical workflow already in place.
To the patient, none of that architecture is visible. A referral sent over WhatsApp gets a reply in their own language within minutes. The system identifies the correct examination, matches it to department pricing, and runs continuously in Chinese or English. AI handles the routine intake and quotation work; every appointment request still surfaces to staff for review before it is confirmed, and every conversation is automatically categorised and tagged so incoming staff can pick up exactly where the AI left off.
Why the Standard AI Playbook Fails in Healthcare
The conventional way to stand up an AI feature — call a hosted model API, wait for a response — does not hold up for a high-volume radiology department processing thousands of patient messages a month, for two reasons.
First, referral letters carry patients' names, dates of birth, and other personal data. Sending that content to an external service outright breaks data protection standards, so de-identification has to happen before anything leaves the hospital's control, not after. Second, real referral letters are full of handwritten annotations, physician shorthand, and institution-specific formats that a general-purpose model reliably misreads. Both constraints point to the same conclusion: the parts of the pipeline that touch raw patient data have to stay inside the hospital's boundary and be tuned to the documents this department actually produces.
Bringing the Model Inside the Walls
Vascue's architecture splits the pipeline along that privacy boundary. Document processing and de-identification, the components that ever touch raw patient data, run on Vascue's own infrastructure inside AWS's ap-east-1 (Hong Kong) region. Identifying details are masked at the boundary before any content is passed onward to downstream AI tasks, and only the data strictly necessary for those tasks makes it through.
AWS Hong Kong Solution Architects worked directly with Vascue's engineering team on the deployment design: network isolation, access controls, compute boundaries, scaling. Local GPU-accelerated compute in the Hong Kong region was a hard requirement, not a nice-to-have — it's what makes keeping the sensitive layer in-region and in-house practical rather than theoretical. What remains of re-identification risk after direct identifiers are stripped is managed through Vascue's technical and security controls: data minimisation, access controls, audit logging, and contractual restrictions on downstream processing. The de-identification layer itself is monitored, logged, and reviewed as part of Vascue's ISO 27001 controls, with a human in that review loop. Both layers are independently audited — AWS infrastructure holds ISO 27001, 27017, 27018, SOC 2, and CSA STAR certifications, and Vascue maintains ISO 27001.
What Becomes Possible
In production, the system has served 6,500+ patients, processed 110,000+ messages, and supported 3,300+ successful bookings. Text-based inquiries typically get a response in 20-30 seconds; image-based referrals, which require extracting content, removing identifying details, matching the examination to department pricing, and generating a quotation, typically resolve in around 60 seconds. WhatsApp support runs 24/7 and handles roughly 46 after-hours inquiries a day, patient demand that previously went unanswered overnight. Staff rate 99.2% of AI responses as correct, under ongoing human review, and around half of patient inquiries now progress to a booked appointment.
The volume this deployment absorbs would previously have required substantial additional headcount. Instead, the existing team gets AI support on the high-volume, repetitive work and keeps its attention on the patient interactions that need clinical judgment. That's the pattern this architecture is built to extend: sensitive processing stays inside Vascue's environment, identifiers get redacted at the boundary, and staff stay in the loop on every clinical decision — in radiology, and in whichever department picks it up next.
Book a demo to see how Vascue's privacy-first architecture maps to your department's workflow.