01 / FEATURE
Radiology triage
Models that flag findings for radiologist review and route urgent cases to the top of the worklist. Radiologist-first workflow — the model never reads alone.
Clinical AI, safely.
Decision support with the clinician in the loop. HSA, MOH and PDPA alignment. Explainability that survives a peer review and a coroner’s enquiry.
The pressure
Why clinical AI is different.
Healthcare AI fails two ways: it never reaches the clinic, or it reaches the clinic and gets quietly switched off. Both happen because the system was built for a benchmark, not a workflow.
Clinical workflows aren’t benchmark datasets. Real radiology, real pathology, real ward rounds — corner cases, low-quality scans, urgent cases jumping the queue. Models trained on Kaggle break on day one.
Documentation burden is crushing. Clinicians spend more time charting than with patients. Every minute saved on documentation is a minute back to clinical care — and burnout reduction.
Multilingual, multi-dialect patient base. Mandarin, Malay, Tamil, Bahasa, Tagalog — and the dialects underneath them. Voice and text systems built for English fail at first contact.
Governance with patient lives in scope. HSA medical-device classification, MOH guidance, PDPA, and clinical governance — explainability isn’t a nice-to-have, it’s how the model survives an MM round.
Where we ship
Use cases we’ve put into production.
Patterns deployed across hospital networks, primary-care groups and pharma in ASEAN — every one with a clinician in the loop and an explainability layer the clinician actually uses.
02 / FEATURE
Pathology decision support
Whole-slide image triage and second-read patterns for cancer screening. Concordance measured against the senior pathologist, not a public dataset.
03 / FEATURE
Clinical documentation summarisation
Discharge summaries, referral letters and progress notes generated from the encounter — clinician edits, signs, owns. Time-on-charting drops materially.
04 / FEATURE
Population-health analytics
Chronic-disease cohort risk stratification — diabetes, CKD, cardiovascular. Care managers see who needs an outreach call this week, with reasons.
05 / FEATURE
Multilingual patient voice agents
Appointment booking, prescription refills, results enquiries — in Mandarin, Malay, Tamil, and the dialects patients actually speak. Hands off to a human the moment a clinical question appears.
06 / FEATURE
Document intelligence for claims
Insurance claims, referral packets, prior-authorisation evidence — extracted with field-level confidence and routed by complexity.
Real-world example
Cutting radiology turnaround by two hours.
A regional hospital network was running a six-hour median turnaround on plain-film chest X-rays — routine cases blocking urgent reads. We deployed a triage model in front of the worklist, reordered by suspected acuity, and gave the radiologist a per-case rationale they could agree, override, or escalate.
Before
Status quo
- 6-hour median turnaround on plain films
- Urgent cases queued behind routine reads
- Radiologists triaging the worklist manually each morning
- No structured rationale on the model output
After
Post-deployment
- Sub-4-hour median; urgent reads under 30 minutes
- Triage automated; radiologist agreed/overrode in two clicks
- Override patterns fed back into weekly retraining
- Every model output ships a clinician-readable rationale
ASEAN regional hospital network · 6 months post-go-live
Solutions that fit
Where to start, by maturity.
AI Sprint — 4 weeks → Validate a clinical use case end-to-end with a working prototype.
Vision AI — imaging → Radiology, pathology, dermatology — clinician-first triage and decision support.
Voice AI — multilingual → Patient voice agents in Mandarin, Malay, Tamil, Bahasa, Tagalog.
Deploy — production SLA → Two to three engineers, full FORGE, monitoring and SLA-backed support.
Compliance & assurance
Frameworks we build against.
Healthcare governance is non-optional. We design for it from the first sprint — not the week before launch.
HSA medical-device alignment. Classification, intended use, risk class and clinical evaluation evidence — designed in from the architecture phase, not retrofitted at submission.
PDPA & MOH data governance. Patient data residency, consent management, and de-identification pipelines built in. Tenant isolation by default; clinical data never leaves the hospital perimeter.
Clinical safety case. Hazard analysis, mitigation, post-market surveillance — the model lifecycle includes the same safety discipline as a medical device, because it is one.
Got a clinical workflow that needs to scale?
Talk to our healthcare lead. We’ll bring the radiology and clinical-documentation case studies, plus the safety-case template.
Talk to healthcare lead
30 minutes · reply within 1 business day