Med-SEAL
A FHIR-native, SEA-LION-powered AI companion for chronic disease care in Southeast Asia.
Med-SEAL is a multilingual, FHIR-native AI companion for chronic disease care, built on SEA-LION, Singapore’s national LLM. It was developed with NUS, Synapxe, IMDA, and MOH Singapore, and aligned with MOH AIHGIe 2.0 guidelines.
Featured as an AI Singapore case study.
The problem
Chronic disease patients in Singapore spend roughly 15 minutes a year with their doctor and 4,300 hours managing their health alone. That gap drives 56% medication non-adherence and an estimated SGD 2.5B in preventable A&E costs.
Architecture
Med-SEAL is a Mixture-of-Agents (MoA) medical AI. Every patient interaction is grounded in a FHIR-native clinical knowledge graph rather than a flat health record. Three specialist agents run in parallel:
- Multilingual reasoning agent (SEA-LION-v4) handles conversation in English, Mandarin, Malay, or Tamil
- Medical reasoning agent (Med-SEAL-V1, an adversarially-trained clinical VLM) checks drug interactions and clinical logic
- Knowledge graph retriever pulls real conditions, medications, and lab trends from the EMR
A SEA-LION aggregator resolves conflicts, enforces safety policy, and returns a single grounded answer. A SEA-Guard safety layer wraps both input and output.
Key components
- Companion mobile app: 24/7 chat, medication tracker, vitals dashboard, appointment booking, and culturally aware dietary coaching (hawker food, festive meals)
- 7-section pre-visit brief auto-generated 24 hours before each appointment, giving clinicians 30 days of adherence, biometric, and patient-reported outcome data
- Tiered nudge engine: gentle reminders for missed doses, next-day clinician flags for concerning trends, immediate alerts for dangerous readings
- FHIR R4 integration with OpenEMR, Medplum, and Epic on FHIR
Results
Med-SEAL-V1 is the first medical VLM in Southeast Asia to combine adversarial training during GRPO reinforcement learning with certified robustness guarantees via randomized smoothing.
| Dimension | Result |
|---|---|
| Adversarial robustness | 0.74% attack success rate on OmniMedVQA under PGD-20, 40x lower than the next-best medical VLM |
| Red-teaming | Zero breaches across 230 attack scenarios: AI Verify 2.0 (11/11), Microsoft PyRIT (26/26 OWASP), NVIDIA Garak (33/33) |
| Quality (DeepEval v3.9.6) | 0.986 faithfulness, 0.940 answer relevancy, 0.933 clinical safety; zero bias, zero toxicity |
My role
I led the AI architecture and trustworthy-AI evaluation. My research on adversarial robustness and certified defenses for vision-language models forms the technical backbone of Med-SEAL-V1.