AI Agents Overview

A patient with diabetes doesn’t need a chatbot. They need someone who remembers their medications, notices when their blood glucose is trending up, gently reminds them about the dose they missed this morning, explains their HbA1c in plain language, and knows that nasi lemak for breakfast means adjusting their advice — all in Mandarin, if that’s what they prefer.

That’s what Med-SEAL’s AI agents do. Seven specialised agents work together — one speaks to the patient, one reasons about their medical data, one watches for danger signs, one knows Southeast Asian food, and one prepares their doctor for the next visit. They are orchestrated by a rule-based router and wrapped in a dual-layer safety guard that ensures no harmful content ever reaches a patient.

Med-SEAL Suite employs a multi-agent AI architecture where specialised agents collaborate to deliver 18 smart healthcare features. The agents are orchestrated by a rule-based router and protected by the SEA-LION Guard safety system.

Agent Roster

ID

Agent

Model

Surface

V1

Med-SEAL V1 (Base Model)

aagdeyogipramana/Med-SEAL-V1 (Qwen3-VL-8B fine-tuned)

Both

A1

Companion Agent

MERaLiON + SEA-LION

Patient app

A2

Clinical Reasoning Agent

Qwen3-VL-8B (Med-SEAL)

Both

A3

Nudge Agent

MERaLiON + rule engine

Patient app

A4

Lifestyle Agent

SEA-LION + nutrition KB

Patient app

A5

Insight Synthesis Agent

Qwen3-VL-8B (Med-SEAL)

Both

A6

Measurement Agent

Analytics engine (no LLM)

Both

SYS

SEA-LION Guard

SEA-LION Guard

System-wide

SYS

smolagents Orchestrator

Rule-based router

System-wide

Note: Med-SEAL V1 (med-r1) is the fine-tuned base model for clinical reasoning. In production, SEA-LION v4-32B (AI Singapore) is used as the clinical reasoning backend. When GPU infrastructure is available, med-r1 can be deployed via vLLM. See Med-SEAL V1 Clinical Agent for details.

Agent Class Diagram

@startuml skinparam classAttributeIconSize 0 hide circle abstract class BaseAgent { agentId : string model : string -- +process(input: AgentInput): AgentOutput +readFHIR(query: FHIRQuery): Bundle } class CompanionAgent { language : string -- +chat(message, patientId): Reply +rephrase(technicalText): string +deliverPRO(questionnaire): Response } class ClinicalReasoningAgent { -- +analyse(conditions, meds, obs): ClinicalResponse +checkInteractions(meds): Warning[] } class NudgeAgent { ruleEngine : RuleEngine -- +evaluate(triggers): Nudge[] +escalate(severity, patientId): Alert } class LifestyleAgent { nutritionKB : KnowledgeBase -- +recommend(diet, conditions): Advice +checkFoodInteractions(meds, food): Warning[] } class InsightSynthesisAgent { lookbackDays : number -- +generateBrief(patientId): Composition +synthesise(adherence, biometrics, PROs): Summary } class MeasurementAgent { -- +computePDC(meds, admin): number +computeTrends(observations): TrendData +generateMeasureReport(): MeasureReport } class SEALIONGuard { -- +checkInput(content): GuardDecision +checkOutput(content): GuardDecision +redactPII(text): string } class Orchestrator { -- +classify(message): Intent +route(intent): BaseAgent } enum GuardDecision { PASS FLAG ESCALATE BLOCK } BaseAgent <|-- CompanionAgent : A1 BaseAgent <|-- ClinicalReasoningAgent : A2 BaseAgent <|-- NudgeAgent : A3 BaseAgent <|-- LifestyleAgent : A4 BaseAgent <|-- InsightSynthesisAgent : A5 BaseAgent <|-- MeasurementAgent : A6 Orchestrator --> BaseAgent : routes to SEALIONGuard --> BaseAgent : wraps SEALIONGuard ..> GuardDecision : returns @enduml

Agent Descriptions

V1 — Med-SEAL V1 (Base Model)

aagdeyogipramana/Med-SEAL-V1 is the fine-tuned base model (Qwen3-VL-8B-Thinking) that serves as the clinical reasoning backbone. It is purely the LLM — not an agent. Agents A2 and A5 call this model for clinical reasoning and pre-visit brief generation respectively.

Not deployed in demo — ChatGPT is used as a substitute. No code changes needed to switch; only environment variables. See Med-SEAL V1 Clinical Agent for full model card and technical details.

A1 - Companion Agent

The patient’s primary conversational interface. Powered by MERaLiON for empathetic, culturally-aware dialogue and SEA-LION for multilingual support (English, Chinese, Malay, Tamil). Delegates complex queries to specialised agents and rephrases technical responses in patient-friendly language.

A2 - Clinical Reasoning Agent

The medical reasoning engine powered by aagdeyogipramana/Med-SEAL-V1 (Qwen3-VL-8B fine-tuned). Reads FHIR clinical data (conditions, medications, observations) and synthesises evidence-based clinical responses. Used by both patient-facing features (via A1) and clinician-facing features.

A3 - Nudge Agent

The proactive engagement engine combining MERaLiON for message generation with a configurable rule engine for trigger evaluation. Handles medication reminders, appointment alerts, biometric threshold warnings, and tiered escalation to clinicians.

A4 - Lifestyle Agent

Specialises in dietary and lifestyle recommendations using SEA-LION with a nutrition knowledge base. Provides culturally-appropriate suggestions grounded in Southeast Asian food (hawker centre dishes, festive food alternatives) with drug-food interaction checking.

A5 - Insight Synthesis Agent

Generates pre-visit clinical briefs for clinicians using Qwen3-VL-8B. Synthesises 30 days of patient-side data (adherence, biometrics, PROs, engagement, flags) into structured FHIR Compositions delivered via CDS Hooks.

A6 - Measurement Agent

A non-LLM analytics engine that computes clinical metrics: medication adherence (PDC), biometric trends, PRO score changes, engagement frequency, readmission rates. Produces FHIR MeasureReports for both individual patients and population cohorts.

Orchestration

The smolagents Orchestrator classifies incoming requests and routes them:

SEA-LION Guard

A system-wide safety layer that wraps all agent interactions:

Input Gate

  • Prompt injection detection

  • Multilingual toxicity filter (EN, ZH, MS, TA)

  • PII redaction

  • FHIR reference validation

Output Gate

  • FHIR $validate (profile conformance)

  • $validate-code (terminology binding)

  • Hallucination check (findings vs source data)

  • Clinical harm filter

Decisions

  • PASS - content delivered as-is

  • FLAG - content delivered with warning annotation

  • ESCALATE - content requires clinician review before delivery

  • BLOCK - content rejected entirely

Feature-to-Agent Matrix

Feature

A1

A2

A3

A4

A5

A6

F01 Chat

PRIMARY

delegate

F02 Medication

display

PRIMARY

alert

compute

F03 Nudge

PRIMARY

F04 PROs

PRIMARY

trigger

consume

F05 Wearables

context

alert

PRIMARY

F06 Clinician summary

PRIMARY

data

F07 Dietary

relay

PRIMARY

F08 Outcome framework

PRIMARY

F09 Dashboard

PRIMARY

F10 Escalation

PRIMARY

enrich

F11 Anticipation

PRIMARY

data

F12 Appointments

PRIMARY

trigger

pre-visit

F13 Caregiver

PRIMARY

forward

data

F14 Education

PRIMARY

trigger

assist

F15 Timeline

PRIMARY

F16 Readmission

flag

consume

PRIMARY

F17 A/B framework

PRIMARY

F18 Satisfaction

PRIMARY

trigger

compute

LLM Configuration

The AI Service connects to SEA-LION v4-32B (AI Singapore) for clinical reasoning and conversation:

# SEA-LION API (primary — powers all agents)
MEDSEAL_SEALION_API_URL=https://api.sea-lion.ai/v1
MEDSEAL_SEALION_API_KEY=<your-key>
MEDSEAL_SEALION_MODEL=aisingapore/Qwen-SEA-LION-v4-32B-IT
MEDSEAL_SEAGUARD_MODEL=aisingapore/SEA-Guard

# Per-agent temperature tuning
MEDSEAL_TEMPERATURE_CLINICAL=0.3
MEDSEAL_TEMPERATURE_COMPANION=0.7
MEDSEAL_TEMPERATURE_NUDGE=0.6
MEDSEAL_TEMPERATURE_LIFESTYLE=0.5
MEDSEAL_TEMPERATURE_INSIGHT=0.2

See Feature Specifications (F01-F18) for the full I/O specification of all 18 features.