Hospital Agentic AI Eco-System — 4 PA · 9 Modul · Your Pace

🎯 Visi: Pesakit Bercakap, Sistem Mendengar, Doktor Bersedia

Selalu kita dengar — pesakit malu cakap depan-depan, atau lupa simptom penting masa konsult. Doktor pula tergesa-gesa, masa konsult 8-10 minit, susah dapat full picture. Farmasi nampak preskripsi tapi tak nampak konteks. Admin uruskan queue secara manual.

Macam mana kalau pesakit boleh mengadu kat AI dulu sambil tunggu giliran, AI siasat full simptom, dan bila pesakit masuk bilik konsult — doktor dah ada briefing lengkap di skrin? Tiada info tercicir, tiada masa terbuang.

🤝

AI Bridge, Bukan AI Diagnosis

AI bukan ganti doktor. AI jambatan antara pesakit yang malu/sibuk dgn doktor yang sibuk. Doktor sentiasa pegang keputusan akhir.

⚡

Wait-Time = Productive Time

10-30 minit pesakit tunggu giliran tu, dia mengadu kat AI. Bila masuk bilik, dah ready. Doktor jimat 2-3 minit per pesakit × 30 pesakit/hari = 1.5 jam.

🔄

Lifecycle Continuity

Pesakit ngadu → konsult → farmasi → ada alergi? Ngadu balik kat AI yang sama. AI ingat sejarah, cross-ref preskripsi, alert doktor + farmasi.

📱 Tiga Cerita Pesakit — Macam Mana Sistem Bantu

Pendekatan paling jujur — tunjuk dari sudut pesakit, bukan dari sudut sistem. Ini yang client mahu nampak.

▸

Cerita 1 — Pak Mat, 45 tahun · Sakit Kepala 3 Hari

Self-care friendly · Bukan kritikal · Doktor optional

1

Pak Mat buka app, taip "sakit kepala dah 3 hari, tak hilang".

Patient PA respond BM. Tanya: lokasi sakit, jenis (berdenyut/ditekan), demam tak, muntah tak, pandangan kabur tak.

2

Pak Mat jawab: "berdenyut sebelah kanan, takde demam, takde muntah".

PA klasifikasi: kemungkinan migraine, BUKAN red flag. Triage = HIJAU (self-care).

3

PA cadang remedi rumah: rehat bilik gelap, hidrasi, paracetamol biasa.

⚠️ "Kalau lebih 24 jam tak baik, atau tambah demam/muntah/kabur, sila ke klinik." Citation: MOH headache CPG.

4

Pak Mat: "Aku rasa nak jumpa doktor jugak."

PA: "OK. Klinik berhampiran ko: Klinik Sihat Sejahtera (1.2km, 3 org tunggu) · Klinik Damai (2.8km, kosong). Mana yang ko pilih?"

5

Pak Mat pilih Klinik Sihat. PA book slot, hantar briefing pre-consult.

Dr. Aminah dapat notif: "Pak Mat (45L) · sakit kepala berdenyut 3 hari · no red flag · suspect migraine · self-care done · masuk dlm 15 minit."

6

Pak Mat masuk bilik. Dr. Aminah dah ready.

"Pak Mat, sakit kepala kanan kan? Berdenyut? Saya nak konfirm beberapa benda..." — terus masuk topik. Konsult 4 minit (vs biasa 8 minit).

Nilai: Pak Mat dapat advice cepat. Klinik dapat lead pre-qualified. Doktor jimat masa. Tiada info tercicir.

▸

Cerita 2 — Mak Cik Salmah, 58 tahun · Sakit Dada Tiba-Tiba

RED FLAG · Auto-escalate · 999 advisory · Hospital routing

1

Mak Cik Salmah buka app: "Sakit dada, sesak nafas, peluh-peluh."

PA detect RED FLAG immediately — sakit dada + sesak + diaphoresis = possible cardiac event.

2

PA cut conversation pendek, paparkan amaran besar:

⚠️ "Mak Cik, ini boleh jadi serangan jantung. Telefon 999 SEKARANG atau minta orang hantar ke ED Hospital terdekat. Jangan pandu sendiri."

3

2 butang besar: [📞 Call 999] [🏥 Hospital Terdekat - ED]

PA tunjuk Hospital A (ED) 4.2km · Hospital B 6.1km. Kongsi lokasi GPS auto bila pesakit tap mana-mana butang.

4

Pesakit tap "Hospital A · ED".

PA hantar pre-arrival packet ke ED triage station: "Pesakit (58F) · sudden chest pain + dyspnea + diaphoresis · ETA via own transport ~12 min · suspect ACS · possible STEMI."

5

ED Hospital terima alert. Cath lab on standby.

Bila Mak Cik tiba, ECG ready, IV line crew ready, cardio on-call notified. Door-to-balloon time potong 15-20 minit.

Nilai: Pesakit kritikal tak terbantut "tunggu giliran". AI auto-escalate dan trigger pre-arrival workflow. Nyawa selamat.

▸

Cerita 3 — Adik Aiman, 24 tahun · Reaksi Alergi Selepas Antibiotik

Lifecycle continuity · Cross-PA loop · Pharmacy + Doctor alerted

1

Aiman dah jumpa doktor 2 hari lalu — preskripsi Amoxicillin untuk URTI.

Patient PA dah simpan: doktor, ubat, dos, tarikh. Doctor PA pula simpan SOAP note + plan.

2

Hari ke-2, Aiman buka app: "Aku gatal-gatal, badan keluar ruam merah."

Patient PA cross-ref: "Ko mula Amoxicillin 2 hari lepas? Ini boleh jadi reaksi alergi ubat."

3

PA tanya 4 soalan: bibir/muka bengkak? susah nafas? pening? nadi laju?

Aiman jawab "tidak" untuk semua. PA: "OK, ini kemungkinan reaksi sederhana. STOP antibiotik. Ambil antihistamine. Ko nak jumpa doktor balik?"

4

PA terus alert 3 pihak parallel:

→ Doctor PA (Dr. Khalid yg preskripsi) — alergi flag added to patient record permanent.
→ Pharmacy PA (klinik tu) — block future Amoxicillin/penicillin order.
→ Admin PA — auto-book follow-up slot esok.

5

Aiman pergi klinik LAIN minggu depan untuk benda lain.

Bila doktor baru nak prescribe — Pharmacy PA alert: "⚠️ Patient ada penicillin allergy (recorded 24 Apr, klinik X)." Auto-cegah preskripsi yg sama.

Nilai: Pesakit yg malu/lupa nak email doktor — boleh ngadu pd PA bila-bila. PA jaga lifecycle, broadcast ke seluruh eco. Adverse event tak terlepas.

👥 Empat AI Personal Assistant — Satu Untuk Setiap Stakeholder

Setiap PA ada peranan, tone, dan keupayaan berbeza. Tapi mereka berkongsi memori — apa pesakit cakap kat Patient PA, doktor nampak (dgn consent).

👤 Patient Personal Assistant

Kawan Ngadu Pesakit

Intake simptom dlm BM/EN rojak
Triage: hijau (self-care) · kuning (klinik) · merah (ED/999)
Cadang remedi rumah dgn citation MOH CPG
Klinik locator + appointment booking
Lifecycle: ngadu → konsult → farmasi → follow-up
Adverse reaction loop (alergi, side effect)

🩺 Doctor Personal Assistant

Co-Pilot Klinikal Doktor

Pre-consult briefing packet sebelum pesakit masuk
Ambient scribe — audio konsult → SOAP note draft
CDSS — DDx, treatment options, citation CPG
Order set drafting (lab, imaging, ubat)
Patient history retrieval — semantic search
Follow-up planner + recall scheduling

💊 Pharmacy Personal Assistant

Penjaga Keselamatan Ubat

DDI check (drug-drug interaction)
Allergy cross-reference (per-pesakit history)
Dose-by-weight (paeds), renal/hepatic adjust
Halal pharma filter (Malaysia context)
Generic substitution dgn cost compare
Adverse reaction intake + back-broadcast

🏢 Admin Personal Assistant

Pengurus Operasi Klinik

Smart queue — priority by triage colour
Appointment booking + WhatsApp reminder
No-show prediction + intervention
Billing automation (panel claim, e-Invoice LHDN)
Daily report — pesakit, revenue, follow-up due
Resource scheduling (doktor, bilik, alat)

Cara mereka bekerjasama: Patient PA jadi orkestrator pesakit — dia panggil Doctor PA bila perlu briefing, Admin PA bila perlu booking, Pharmacy PA bila ada preskripsi. Setiap PA log activity ke audit trail bersama. Pesakit dgn pesakit punya data asing, tiada cross-leak.

🧩 Sembilan Modul — Pilih Ikut Kepentingan Klinik

Bukan all-or-nothing. Setiap modul boleh deploy standalone atau compose dgn modul lain. Klinik solo boleh start dgn 1-2 modul, tambah bila ready.

✓ STANDALONE — berfungsi sendiri ✗ DEPENDENT — perlu modul prasyarat ⊕ CROSS-CUTTING — always-on, semua modul guna

M1 · PSPA ✓ STANDALONE

Patient Symptom PA

Pesakit ngadu via app/WhatsApp. AI intake simptom, triage 3-warna, cadang self-care atau escalate.

Standalone value: Klinik dapat WhatsApp triage bot — kurangkan walk-in tak perlu, tingkatkan walk-in pre-qualified.

Compose dgn: M2 · M3 · M6

M2 · PCBP ✗ NEEDS M1

Pre-Consult Briefing Packet

Bridge dari Patient PA ke skrin doktor. Doktor dapat ringkasan simptom + triage + history sebelum pesakit masuk bilik.

Standalone value: Doktor jimat 2-3 minit per konsult. 30 pesakit/hari = 1.5 jam jimat.

Compose dgn: M1 (mandatory) · M4

M3 · CLOC ✓ STANDALONE

Clinic Locator & Smart Routing

GPS pesakit → senarai klinik berhampiran dlm rangkaian, dgn queue length, jarak, dan masa anggaran.

Standalone value: Klinik dapat lead routing — pesakit yg sesuai datang, queue dipancut.

Compose dgn: M1 · M7 · M8

M4 · DRPA ✓ STANDALONE

Doctor PA — Clinical Co-Pilot

Ambient scribe (audio→SOAP), CDSS, order set drafting, patient history search. Boleh guna tanpa Patient PA.

Standalone value: Mana-mana doktor — solo GP, hospital ward, klinik pakar — dapat AI scribe + decision support.

Compose dgn: M2 (lebih powerful) · M5

M5 · PHPA ✓ STANDALONE

Pharmacy PA

DDI checker, allergy cross-ref, dose-by-weight, halal filter, generic substitution. Per-pesakit memory bila ada M1.

Standalone value: Mana-mana farmasi — klinik in-house atau community pharmacy — dapat safety net automatik.

Compose dgn: M4 · M6

M6 · ARXL ✗ NEEDS M1+M5

Adverse Reaction Loop

Pesakit ngadu side effect via Patient PA → cross-ref preskripsi terkini → alert doktor + farmasi → block re-prescribe.

Standalone value: Permanen patient allergy/intolerance registry — follow pesakit ke mana-mana klinik dlm rangkaian.

Compose dgn: M1+M5 (mandatory) · M8

M7 · ADPA ✓ STANDALONE

Admin PA — Operasi Klinik

Smart queue, appointment, WAHA reminder, no-show predict, billing, panel claim, e-Invoice LHDN, daily report.

Standalone value: Klinik kurangkan staf admin/receptionist, atau tingkatkan kapasiti tanpa hire.

Compose dgn: M1 · M2 · M3

M8 · XCCN ✗ NETWORK

Cross-Clinic Continuity

Pesakit data follow pesakit (dgn consent) merentas klinik dlm rangkaian. Allergi, kronik, history — sentiasa available.

Standalone value: Bukan sekadar feature — ini network effect. Lebih ramai klinik join, lebih bernilai untuk pesakit.

Compose dgn: M1+ (mandatory) · M6

M9 · AUCM ⊕ CROSS-CUTTING

Audit · Compliance · Guardrails

PDPA filter, hallucination guard, HITL approval gate, MOH audit trail. Sentiasa ON kalau ada modul produksi.

Standalone value: Tiada — ini foundational layer. Mandatory untuk apa-apa deployment yg sentuh data klinikal.

Compose dgn: Semua modul (mandatory)

⚠️ Hard rule: M9 wajib bundle dgn mana-mana modul produksi yang sentuh data pesakit. Tiada negotiation — ini PDPA + clinical safety baseline.

🚪 Tiga Adoption Paths — Pilih Yang Sesuai

Tiga jenis client biasa. Setiap satu ada combo modul yang masuk akal untuk start.

PATH A · STARTER

Klinik Solo

GP solo, klinik panel, klinik pakar single doktor

Start dgn:

M1 · Patient Symptom PA (WhatsApp triage)

M4 · Doctor PA (ambient scribe)

M9 · Audit (mandatory)

Tambah bila ready:

+ M2 (Pre-consult briefing)

+ M5 (Pharmacy in-house)

+ M7 (Admin operasi)

PATH B · GROWTH

Group Klinik / Chain

Rangkaian klinik 5-50 cawangan, panel insurance, group pakar

Start dgn:

M1 · Patient Symptom PA

M3 · Clinic Locator (network advantage)

M7 · Admin PA (centralized ops)

M8 · Cross-Clinic Continuity

M9 · Audit

Tambah bila ready:

+ M2 + M4 (per-cawangan)

+ M5 + M6 (pharmacy chain)

PATH C · ENTERPRISE

Hospital / Hospital Group

Hospital pakar, hospital korporat, hospital kerajaan

Full eco:

M1 · Patient PA (community channel)

M2 · Pre-consult briefing

M4 · Doctor PA (hospital-wide)

M5 + M6 · Pharmacy + Adverse loop

M7 · Admin PA

M8 · Cross-clinic continuity

M9 · Audit + on-prem option

M3 optional — hospital biasanya destination, bukan locator.

🔗 Composability Matrix — Apa Berfungsi Dgn Apa

Cara baca: row = modul yg ada, column = modul yg compose. Hijau = unlock value baru, merah = perlu prerequisite, hitam = self.

M1

M2

M3

M4

M5

M6

M7

M8

M9

M1

●

unlocks

+ctx

unlocks

+ctx

unlocks

+ctx

M2

needs

●

+ctx

M3

+ctx

●

+ctx

M4

+ctx

●

+ctx

M5

+ctx

●

unlocks

+ctx

M6

needs

+ctx

needs

●

+ctx

M7

+ctx

●

+ctx

M8

needs

+ctx

●

+ctx

M9

guards

●

"unlocks" = capability baru terbuka. "+ctx" = modul tu kerja lagi baik dgn konteks tambahan. "needs" = prerequisite mandatory. "guards" = M9 layer keselamatan untuk semua.

🏗️ Modular Architecture — Plug-In Design

Setiap modul = satu service independen, communicate via event bus + tool registry. Tambah/buang modul tak break yang lain.

┌──────────────────────────────────────────────────────────────────────┐
│  CLIENT INTERFACES                                                   │
│  ▸ Patient app (mobile/PWA)  ▸ Doctor desktop  ▸ Pharmacy POS  ▸ Admin
└──────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│  M9 · GUARDRAILS LAYER (cross-cutting)                               │
│  PDPA filter · HITL gate · Hallucination guard · Audit logger        │
└──────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│  EVENT BUS + ORCHESTRATOR                                            │
│  Patient.symptom.intake  ·  Patient.escalate  ·  Doctor.briefing.ready
│  Pharmacy.dispense  ·  Adverse.reaction.report  ·  Appointment.booked│
└──────────────────────────────────────────────────────────────────────┘
        │           │           │           │           │
        ▼           ▼           ▼           ▼           ▼
   ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐
   │M1 PSPA │  │M4 DRPA │  │M5 PHPA │  │M7 ADPA │  │ ...    │
   │Patient │  │Doctor  │  │Pharmacy│  │Admin   │  │ M2/3/6 │
   │  PA    │  │  PA    │  │  PA    │  │  PA    │  │ /8 etc │
   └────────┘  └────────┘  └────────┘  └────────┘  └────────┘
        │           │           │           │
        └───────────┴───────────┴───────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│  mcp-medix · TOOL REGISTRY                                           │
│  clinical/* · medication/* · patient/* · imaging/* · billing/*       │
│  comms/* · scheduling/* · admin/*                                    │
└──────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌──────────────────────────────────────────────────────────────────────┐
│  FOUNDATION (existing)                                               │
│  medix.alesa.my (HIS) · medic.alesa.my (6 AI services LIVE)          │
│  + FHIR R4 · NPRA Drug DB · MOH CPG library                          │
└──────────────────────────────────────────────────────────────────────┘

Plug-in Design

Modul bercakap melalui event bus standard. Tambah modul baru = subscribe event yang relevan. Tiada redeploy.

Per-Tenant Isolation

Setiap klinik/hospital ada tenant key tersendiri. Data tak bocor. M8 buat sharing dgn consent eksplisit.

Foundation Reuse

Modul tak rebuild dari sifar — guna 6 AI services medic.alesa.my + 86 jadual medix.alesa.my sebagai base.

🛡️ M9 Guardrails — Lima Lapisan Keselamatan

Wajib ON untuk semua deployment produksi. Healthcare = nyawa, tiada exception.

🔴 1. Clinical Safety Gate

Tindakan high-risk (controlled drug, surgery, paeds dose >2× standard, transfusion) WAJIB consultant approval. Default OFF — owner enable per-modul.

🟠 2. PDPA Filter

Semua input ke LLM lalu PII stripper dulu. IC, nama penuh, alamat ditukar token. Output detok di UI sahaja. Coverage 12 PII types.

🟣 3. Hallucination Guard

Setiap dx/dose/cadangan WAJIB cite source (NPRA, MOH CPG, WHO, hospital SOP). Tiada source = tiada paparan.

🔵 4. Human-in-the-Loop

PA tidak execute — mereka propose. Doktor/farmasis tap approve sebelum tindakan jalan. Timeout default reject.

🟢 5. Audit & Observability

Setiap tool call, LLM call, HITL approve — log lengkap, queryable. MOH-ready audit trail. 7-tahun retention.

⚙️ Tech Stack

Foundation (existing · LIVE)

Laravel 11 Filament 3 MariaDB LiteSpeed

Agent Runtime

Python 3.12 FastAPI LangGraph MCP SDK Pydantic Celery

LLM Options

OpenAI gpt-4o-mini Llama 3.1 70B (on-prem) Qwen 2.5-Med (on-prem) BGE-M3 embed

Memory & State

Redis (working) pgvector (long-term) PostgreSQL (state)

Comms & Integration

FHIR R4 HL7 v2 DICOMweb WAHA WhatsApp e-Invoice LHDN

Observability & Security

OpenTelemetry Loki + Grafana Sentry Vault PII Detector

🖥️ On-Prem AI Infrastructure — Bukti Kapasiti 200 Klinik

Bukan cloud rental yang berkemungkinan ditarik balik. Bukan API kotak hitam. Fizikal AI server di rak datacenter Malaysia — ALESA pegang hardware, klien pegang data.

Sizing di bawah berdasarkan andaian konservatif: 200 klinik × max 500 pesakit/klinik/bulan = 100,000 pesakit/bulan eco-sistem. Senario dipilih untuk peak load Monday/cuti umum.

📊 SIZING FIGURE — 200 KLINIK · 100,000 PESAKIT/BULAN · ON-PREM AI

⓵ FIZIKAL AI RACK

Hardware Specification

TIER-1 INFERENCE (heavy)

2× NVIDIA H100 80GB · Llama 3.1 70B Q5 · CDSS, DDx, complex reasoning

TIER-2 INFERENCE (fast)

2× NVIDIA RTX A6000 48GB · Llama 3.1 8B FP16 · Triage, intake, classification

EMBEDDING + ASR

2× RTX 4090 24GB · BGE-M3 (multilingual embed) + Whisper-Large v3 (BM/EN ASR)

DATABASE CLUSTER

2× Dell R7615 · 96-core EPYC · 512GB RAM · 30TB NVMe RAID-10 · primary + replica

STORAGE + NETWORK

100TB Ceph object storage · 25Gbps backbone · dual UPS 30kW · redundant L3 switch

Total VRAM: ~336GB · CPU cores: 192 · RAM: 1TB · Storage: 130TB

⓶ KAPASITI MATH

200 Klinik → Beban Sebenar

Klinik dlm rangkaian 200

×

Max pesakit/klinik/bulan ~500

=

Total/bulan 100,000

÷ 30

Avg/hari ~3,300

peak ×1.5

Peak day (Isnin) ~5,000

Peak hour ~375

Concurrent sessions peak ~25-40

Token workload/bulan

100K pesakit × ~17,500 tokens = ~1.75 BILLION tokens/bulan

⓷ THROUGHPUT & SLA

Performance Target

PEAK REQUEST RATE

6 req/min sustained

AGGREGATE INFERENCE

~750 tok/sec

LATENCY p50 / p99

1.2s / 4.5s end-to-end

CONCURRENT CAPACITY

100+ sessions

UPTIME SLA

99.5% target ops

HEADROOM

Spec ini handle ~3-4× current load — boleh scale ke 500-700 klinik (atau bila avg pesakit naik dari 500 ke 1,000+/bulan) tanpa hardware tambahan dgn smart-routing.

⓵ FIZIKAL rack di datacenter Malaysia ⓶ KAPASITI 200 klinik · 100K pesakit/bulan ⓷ THROUGHPUT 750 tok/s · 99.5% SLA Sizing v1.1 · 26 Apr 2026 · 500 pesakit/klinik baseline

💰 On-Prem vs Cloud LLM — Kos/Bulan (200 klinik baseline)

Cloud LLM (gpt-4o-mini)

1.75B tokens × public pricing

~RM 450K

On-Prem (ALESA Rack)

Colo + power + maintenance

~RM 50K

Kos/pesakit on-prem: ~RM 0.50 · Cloud: ~RM 4.50 · ~9× lebih murah

⚠️ Capex hardware ~RM 800K-1.2M (one-time). ROI ~14-18 bulan pd 500/klinik baseline; lebih cepat bila volume tumbuh ke 1K+/klinik.

🏗️ Deployment Options

Option A · ALESA Datacenter (Default)

Klinik subscribe SaaS. Inference jalan di rak ALESA. Klien jimat capex, jaga OpEx je. Sesuai untuk Path A & B.

Option B · Klien On-Site (Hospital)

Hospital pakar/korporat ada datacenter sendiri — rak diletak di premis. Data tak keluar premis. Sesuai untuk Path C SULIT.

Option C · Hybrid

Inference on-prem, embedding/cache di cloud. ALESA urus ops, klien akses via VPN. Compromise sesuai untuk group klinik.

Option D · Cloud Burst (Peak only)

Default on-prem. Bila peak overflow, route ke cloud LLM dgn PII-strip. Cost-efficient untuk spike events.

🎯 Kenapa "Fizikal AI Server" Penting Untuk Pitch

PDPA & Sovereignty

Data klinikal Malaysia tinggal di Malaysia. Tiada cross-border concern. MOH audit boleh masuk premis fizikal.

Kos Stabil

Tiada surprise bill bila API provider naikkan harga atau tarik service. Capex satu kali, OpEx predictable.

Bukti Komitmen

Hardware fizikal = ALESA own infrastructure. Bukan reseller cloud. Klien percaya kami serious untuk healthcare jangka panjang.

🎚️ Tiga Tier Server — Pilih Ikut Skala

Tiada one-size-fits-all. Low-End untuk pilot/PoC. Mid-End untuk klinik tumbuh. Hi-End untuk full eco scale.

TIER 0 · LOW-END

RM 80-150K capex

~5-30 klinik · ~2.5-15K pesakit/bulan

INFERENCE GPU

1× RTX 4090 24GB (24GB VRAM)
Llama 3.1 8B FP16 + multiplexed BGE-M3 + Whisper

COMPUTE NODE

Tower workstation: AMD Threadripper PRO 7965WX (24-core) · 128GB DDR5 · 8TB NVMe Gen4

STORAGE + NETWORK

25TB NAS NVMe · 1Gbps · single UPS 5kW · standalone deployment

PERFORMANCE

~250

tok/sec

~15

concurrent

98.0%

SLA

OpEx anggaran

~RM 8-15K/bulan · capex ROI ~10-12 bulan

Sesuai untuk: Pilot/PoC · klinik solo · dev/staging environment · Doc Zam demo deployment

TIER 1 · MID-END

RM 350K-450K capex

~30-150 klinik · ~15-75K pesakit/bulan

INFERENCE GPU

2× NVIDIA L40S 48GB (96GB VRAM)
Llama 3.1 70B Q5 + 8B FP16 fallback

EMBED + ASR

1× RTX 4090 24GB · BGE-M3 + Whisper-Large v3 (multiplex)

DATABASE

1× Dell R7615 · 64-core EPYC · 256GB RAM · 16TB NVMe RAID-10
+ 1× R660 backup replica

STORAGE + NETWORK

50TB Ceph (2-replica) · 10Gbps · single UPS 15kW · L3 switch redundant

PERFORMANCE

600

tok/sec

~40

concurrent

99.0%

SLA

OpEx anggaran

~RM 25-30K/bulan · capex ROI ~12-14 bulan

Sesuai untuk: Path A (klinik solo cluster) · Path B awal (group klinik 5-50 cawangan) · pilot fasa pertama

TIER 2 · HI-END

RM 1.5M-2.0M capex

~150-500 klinik · ~75-250K pesakit/bulan

TIER-1 INFERENCE (heavy)

4× NVIDIA H100 80GB (320GB VRAM)
Llama 3.1 70B Q5 + room untuk 405B Q4 masa depan

TIER-2 INFERENCE (fast)

4× NVIDIA RTX A6000 48GB (192GB VRAM)
Llama 3.1 8B FP16 fleet untuk triage burst

EMBED + ASR (redundant)

2× RTX 4090 24GB · BGE-M3 + Whisper Large v3 dgn HA failover

DATABASE CLUSTER (HA)

2× Dell R7615 (primary+replica) · 96-core EPYC · 512GB RAM · 30TB NVMe RAID-10
+ 1× R760xs (audit log dedicated)

STORAGE + NETWORK

200TB Ceph (3-replica) · 25Gbps backbone · dual UPS 30kW · redundant cooling 25kW

PERFORMANCE

2,500

tok/sec

~150

concurrent

99.5%

SLA

OpEx anggaran

~RM 55-70K/bulan · capex ROI ~15-18 bulan

Sesuai untuk: Path B mature (group 50+ cawangan) · Path C (hospital korporat / hospital group) · network-effect scale

Metric

LOW-END

MID-END

HI-END

Nota

Klinik dilayan

5-30

30-150

150-500

Max 500 pesakit/klinik/bulan

Pesakit/bulan

~2.5-15K

~15-75K

75-250K

Termasuk follow-up + alergi

GPU / VRAM

1× 4090 / 24GB

3 GPU / 120GB

10 GPU / 512GB

RTX 4090 → L40S → H100+A6000

Throughput

~250 tok/s

~600 tok/s

~2,500 tok/s

Aggregate sustained

Concurrent peak

~15

~40

~150

Active sessions simultaneous

Uptime SLA

98.0%

99.0%

99.5%

HA + redundant power di Mid/Hi

Capex (one-time)

RM 80-150K

RM 350-450K

RM 1.5-2.0M

Anggaran 2026 GPU pricing

OpEx (bulanan)

RM 8-15K

RM 25-30K

RM 55-70K

Colo + power + maintenance

Kos/pesakit

~RM 1.30

~RM 0.60

~RM 0.40

Lebih besar = lebih murah

📌 Upgrade path: Low-End → Mid-End (~RM 250-300K tambah · 6-bulan ROI) → Hi-End (~RM 1-1.2M tambah · scale-out tanpa downtime). Mula kecik, tumbuh ikut load sebenar.

💻 Sisi Klinik — Pakai Hardware Sedia Ada

95% AI compute jalan di rak ALESA. Klinik tak perlu beli laptop atau PC baru. Hardware sedia ada di kaunter, bilik konsult, dan farmasi cukup — kerana semua interaksi melalui browser PWA.

⚡ ARCHITECTURE PRINCIPLE

Browser-First · PWA · Zero-Install

Setiap PA (Patient · Doctor · Pharmacy · Admin) = Progressive Web App. Buka Chrome, login, kerja. Heavy AI inference — semua di server ALESA. Hardware klinik hanya jadi UI terminal + capture peripheral (mic, camera, scanner).

KAUNTER · ADMIN

Receptionist / Counter Staff

🏢

Workload: queue, booking, register, billing, panel claim

MIN SPEC

Intel i3 8th gen / Ryzen 3
4GB RAM / SSD 128GB
Chrome 110+

PERIPHERAL

Receipt printer
Barcode scanner
QR cam (existing)

✓ Hampir semua kaunter klinik dah ada hardware ni

BILIK KONSULT · DOKTOR

Doctor (Standard Mode)

🩺

Workload: briefing review, SOAP edit, order set, e-prescription

MIN SPEC

Intel i5 10th gen / Ryzen 5
8GB RAM / SSD 256GB
Chrome 110+ / mic built-in

PERIPHERAL

Mic (laptop built-in OK)
Webcam (optional)
Tablet stylus (optional)

✓ Audio capture via WebRTC ke server — tiada local AI

BILIK KONSULT · PRO MODE

Doctor (Offline-Capable)

⚡

Workload: sama + ambient ASR offline (jika network down)

MIN SPEC

Intel i7 / Ryzen 7
16GB RAM / SSD 512GB
iGPU OK (Iris Xe / Radeon)

PERIPHERAL

Quality lapel mic
Headset (optional)
External webcam

✓ Whisper.cpp local fallback (BM/EN) bila network down

FARMASI · DISPENSING

Pharmacy Counter

💊

Workload: DDI alert, dispense verify, label print, stock

MIN SPEC

Intel i3 8th gen / Ryzen 3
4-8GB RAM / SSD
Chrome 110+

PERIPHERAL

Barcode scanner
Label printer (Zebra etc)
Cash drawer

✓ Existing pharmacy POS hardware compatible

🌐 Bandwidth Klinik

Idle PWA~50 Kbps

Active konsult (text)~200 Kbps

Ambient audio stream~512 Kbps

Image upload (lab/imaging)~2 Mbps burst

Pilih Modul. Mula Kecik. Tumbuh Ikut Pace.

Eco-sistem holistik — tapi adoption modular. Klinik solo boleh start dgn M1+M9. Group klinik scale ke M3+M7+M8. Hospital ambil full eco. Setiap fasa standalone valuable, tiada lock-in.

PATH A · STARTER

M1 + M4 + M9

Klinik solo · Triage bot + AI scribe

PATH B · GROWTH

M1 + M3 + M7 + M8 + M9

Group klinik · Network advantage

PATH C · ENTERPRISE

M1-M9 (full eco)

Hospital · Lifecycle complete

→ medix.alesa.my (HIS) → medic.alesa.my (AI) → Doc Zam Analisis → Lebih Baik Dari Adeahub

Satu Visi Holistik. Sembilan Modul. Ko Pilih Pace.

Pegang. Geser. Lihat.

Berapa cawangan klinik anda?

Apa bajet capex permulaan anda?

Apa priority utama anda?

Sumber IT klinik anda?

🎯 Visi: Pesakit Bercakap, Sistem Mendengar, Doktor Bersedia

📱 Tiga Cerita Pesakit — Macam Mana Sistem Bantu

👥 Empat AI Personal Assistant — Satu Untuk Setiap Stakeholder

Kawan Ngadu Pesakit

Co-Pilot Klinikal Doktor

Penjaga Keselamatan Ubat

Pengurus Operasi Klinik

🧩 Sembilan Modul — Pilih Ikut Kepentingan Klinik

🚪 Tiga Adoption Paths — Pilih Yang Sesuai

🔗 Composability Matrix — Apa Berfungsi Dgn Apa

🏗️ Modular Architecture — Plug-In Design

🛡️ M9 Guardrails — Lima Lapisan Keselamatan

⚙️ Tech Stack

🖥️ On-Prem AI Infrastructure — Bukti Kapasiti 200 Klinik

🎚️ Tiga Tier Server — Pilih Ikut Skala

💻 Sisi Klinik — Pakai Hardware Sedia Ada

Pilih Modul. Mula Kecik. Tumbuh Ikut Pace.