VPN-Only vLLM Infrastructure
AI That Runs Behind a Locked Network
Designated VPS nodes connect to AINNA's central vLLM server through encrypted VPN tunnels. Public internet has no route, no endpoint, and no direct access to the LLM layer.
🔒 Central vLLM · Whitelisted VPS only · Zero public inference ports · Sovereign AI operations
NeuralOps is not a chatbot wrapper. It is a controlled AI operations stack where agents run on isolated VPS nodes, inference happens through private VPN, and business data stays inside your governed environment.
How Users Reach AI Without Exposing vLLM
The user never connects to the vLLM server directly. Requests go to the designated VPS first, then the VPS calls the central vLLM through a private WireGuard VPN tunnel.
Designated VPS Node
Your business app, agents, dashboards, and automation run on an isolated VPS. This VPS becomes the only approved execution node for your AI workflow.
Private VPN Route
AINNA creates a WireGuard tunnel between your VPS and the central vLLM server. The VPS IP and tunnel credentials are whitelisted. Public traffic is rejected by design.
Controlled vLLM Inference
Agents send prompts through VPN, vLLM returns model output, and operations continue on the VPS. No public API keys, no exposed model backend, and no direct GPU access.
AINNA NeuralOps Ecosystem Components
Complete AI infrastructure stack from local models to autonomous agents.
Local LLM Server
Private AI infrastructure running on-premise
Private LLM Infrastructure
Enterprise-grade security and data sovereignty
Agentic AI
Autonomous agents that execute complex workflows
AI Agent Builder
Design and deploy custom AI agents
Local LLM Server
Private AI infrastructure running on-premise
Private LLM Infrastructure
Enterprise-grade security and data sovereignty
Agentic AI
Autonomous agents that execute complex workflows
AI Agent Builder
Design and deploy custom AI agents
NeuralOps Service Plans
Enterprise-grade AI infrastructure with VPN-only security. No public API. No data leakage.
For small businesses and startups
- ✅ Shared VPS instance (isolated multi-tenant)
- ✅ VPN tunnel to central vLLM server
- ✅ Access to 2 LLM models
- ✅ 100K tokens/day
- ✅ Standard support (ticket/email)
- ✅ 3-4 day onboarding
For growing companies and e-commerce
- ✅ Dedicated VPS instance (single-tenant)
- ✅ Dedicated VPN tunnel to vLLM server
- ✅ Access to 5 LLM models
- ✅ 500K tokens/day
- ✅ OpenClaw agent deployment included
- ✅ Priority support (chat + ticket)
- ✅ 5-7 day onboarding
For enterprises and high-security needs
- ✅ Multiple dedicated VPS instances
- ✅ Multiple VPN tunnels (redundancy)
- ✅ Access to ALL 7 LLM models
- ✅ Custom rate limits
- ✅ Full Hermes integration
- ✅ Custom Detached System workflows
- ✅ Dedicated account engineer
- ✅ 99.9% uptime SLA
NeuralOps Service Model — Super-Secured vLLM Infrastructure
Not a private LLM per customer. One centralized, super-secured vLLM server. Only authorized VPS instances connect via VPN. Zero open connections. Hub-and-spoke architecture.
NOT Private LLM Per Customer
No per-client model isolation. Instead, a single optimized vLLM server shared via secure VPN tunnels — more efficient and equally secure.
ONE Centralized vLLM Server
Single server running 7 LLM models. Fully utilized GPU resources. Centralized monitoring, updates, and security patches.
Authorized VPS Only (VPN)
Only designated VPS instances with whitelisted VPN credentials can connect. No public API keys. No exposed endpoints.
Zero Open Connections
No open inference ports. No public admin panel. No exposed model backend. No direct GPU access from outside.
Hub-and-Spoke Topology
Central vLLM server connects to isolated VPS spokes via VPN. If one VPS fails, others remain unaffected. LLM server stays protected.
Audit Trail & Monitoring
Every inference request logged. Centralized monitoring across all VPS connections. Complete auditability for compliance.
NOT Private LLM Per Customer
No per-client model isolation. Instead, a single optimized vLLM server shared via secure VPN tunnels — more efficient and equally secure.
ONE Centralized vLLM Server
Single server running 7 LLM models. Fully utilized GPU resources. Centralized monitoring, updates, and security patches.
Authorized VPS Only (VPN)
Only designated VPS instances with whitelisted VPN credentials can connect. No public API keys. No exposed endpoints.
Zero Open Connections
No open inference ports. No public admin panel. No exposed model backend. No direct GPU access from outside.
Hub-and-Spoke Topology
Central vLLM server connects to isolated VPS spokes via VPN. If one VPS fails, others remain unaffected. LLM server stays protected.
Audit Trail & Monitoring
Every inference request logged. Centralized monitoring across all VPS connections. Complete auditability for compliance.
Data Sovereignty & Security Policy
Your data never leaves your controlled environment. All inference happens behind VPN. No third-party API calls. No data leakage to public models. Security policy (kebijakan keselamatan) enforced at network perimeter.
Your VPS
Business data stays here. Agents process, agents execute. No data leaves your isolated zone.
VPN Tunnel (Encrypted)
Only prompt text travels through. Fully encrypted. No stored logs of your data.
LLM Server
Receives prompt → returns output. No data stored. No data logged. No data shared.
Transform Raw Data Into Structured Intelligence
AINNA transforms raw operational data into structured intelligence. Sales, inventory, product listing, ads, logistics, customer behavior, and financial signals can be converted into dashboards, forecasting models, anomaly detection, and decision support systems.
Data Intelligence
Extract insights from operational data
RAG Knowledge Base
Retrieval-augmented generation with your data
Vector Database
High-performance semantic search and retrieval
Detached Business System
AI-built systems that run independently
Data Intelligence
Extract insights from operational data
RAG Knowledge Base
Retrieval-augmented generation with your data
Vector Database
High-performance semantic search and retrieval
Detached Business System
AI-built systems that run independently
Intelligent Orchestration Across Your Business
AINNA automates repetitive business processes across ecommerce, inventory, reporting, listing management, operations, and internal workflows. The goal is not just automation, but intelligent orchestration between data, rules, people, and systems.
Ecommerce Automation
End-to-end online store management
Inventory Dashboard
Real-time stock tracking and alerts
Sales Report System
Automated revenue analytics and insights
Predictive Analytics
ML-powered forecasting and trends
Ecommerce Automation
End-to-end online store management
Inventory Dashboard
Real-time stock tracking and alerts
Sales Report System
Automated revenue analytics and insights
Predictive Analytics
ML-powered forecasting and trends
Bridging AI Ambition and Real-World Execution
Instead of relying only on chatbot interfaces, AINNA uses AI as a system builder — designing, generating, repairing, and optimizing business applications that continue to operate independently after deployment.
Purpose-Built AI
Every system is designed for specific business outcomes, not generic conversations.
Knowledge Stays Local
Your business logic, data, and insights remain under your control — no external dependencies.
Independent Operation
Systems continue running after deployment, requiring minimal maintenance and zero token costs.
Purpose-Built AI
Every system is designed for specific business outcomes, not generic conversations.
Knowledge Stays Local
Your business logic, data, and insights remain under your control — no external dependencies.
Independent Operation
Systems continue running after deployment, requiring minimal maintenance and zero token costs.
AI Agents That Build Real Systems
AINNA develops AI Agents that assist in system planning, code generation, workflow design, data mapping, error detection, documentation, optimization, and system repair. These agents are not uncontrolled bots. They operate within clear business rules, human approval layers, and defined system boundaries.
System Planning & Design
Agents analyze business requirements and architect optimal system structures.
Code Generation & Development
AI-assisted development produces clean, production-ready code.
Error Detection & Repair
Agents identify bugs, edge cases, and performance issues — then fix them.
Governed Operations
Every agent action follows business rules with human approval and audit trails.
System Planning & Design
Agents analyze business requirements and architect optimal system structures.
Code Generation & Development
AI-assisted development produces clean, production-ready code.
Error Detection & Repair
Agents identify bugs, edge cases, and performance issues — then fix them.
Governed Operations
Every agent action follows business rules with human approval and audit trails.
Central vLLM Server — Locked Behind VPN
The vLLM server is the private inference layer. It runs centrally for performance and model efficiency, but it is not visible to the public internet. Only designated VPS nodes can call it through encrypted VPN.
🔒 No public IP · No public inference port · No public admin panel · No direct GPU access · Whitelisted VPS only.
VPN-Secured vLLM
The vLLM server is reachable only from approved VPS nodes through private WireGuard tunnels.
No Public LLM Backend
No browser, attacker, scraper, or random API client can reach the model backend directly.
Agent Execution Layer
OpenClaw, OpenCode, Detached Systems, and Hermes workflows run on VPS, not on the GPU server.
7-Model Capacity
Central vLLM can serve multiple models while keeping access controlled through VPN-only routing.
VPN-Secured vLLM
The vLLM server is reachable only from approved VPS nodes through private WireGuard tunnels.
No Public LLM Backend
No browser, attacker, scraper, or random API client can reach the model backend directly.
Agent Execution Layer
OpenClaw, OpenCode, Detached Systems, and Hermes workflows run on VPS, not on the GPU server.
7-Model Capacity
Central vLLM can serve multiple models while keeping access controlled through VPN-only routing.
Systems That Run Independently After Creation
Unlike AI tools that require constant prompting, AINNA builds detached systems that run independently after creation.
AI Designs
Agents analyze requirements and architect the optimal system structure.
AI Develops
Code, database schemas, APIs, and interfaces are generated with documentation.
AI Optimizes
Performance tuning, bug fixes, and edge case handling through iterative refinement.
System Runs
Deployed application operates independently — standard web app with automation.
AI Designs
Agents analyze requirements and architect the optimal system structure.
AI Develops
Code, database schemas, APIs, and interfaces are generated with documentation.
AI Optimizes
Performance tuning, bug fixes, and edge case handling through iterative refinement.
System Runs
Deployed application operates independently — standard web app with automation.
Control-First AI Adoption
AINNA's model is control-first. Human-in-the-loop approval, audit trail, system logging, access control, role-based permissions, monitoring, observability, and business rule enforcement are included to keep AI adoption safe, explainable, and manageable.
Governance Framework
Human-in-the-loop approval, audit trail, system logging, access control, and business rule enforcement.
Safe & Explainable AI
Every automated decision can be traced and explained with clear escalation paths.
Human-in-the-Loop
Approval workflows for critical decisions with configurable risk levels.
Monitoring & Audit
Complete logging, observability dashboards, and compliance reviews.
Sovereign Data & Kebijakan
Your data never leaves your controlled VPS environment. Security policy enforced at network perimeter. No third-party exposure.
Safe & Explainable AI
Every automated decision can be traced and explained with clear escalation paths.
Human-in-the-Loop
Approval workflows for critical decisions with configurable risk levels.
Monitoring & Audit
Complete logging, observability dashboards, and compliance reviews.
Real Operational Experience at Scale
AINNA has real operational experience managing large-scale ecommerce workflows with multiple stores, high SKU volume, and automation.
Multi-Store Operations
12 stores across Shopee, Lazada, TikTok Shop, and Shopify with unified inventory sync.
High SKU Volume
45,000+ SKUs managed with variant support, bulk price updates, and automated stock alerts.
Sales Reporting & Analytics
Real-time revenue tracking (RM2.3M GMV/mo), profit margin analysis by channel, and performance insights.
End-to-End Automation
From product listing to order fulfillment — automated workflows that reduce manual work by 68%.
Multi-Store Operations
12 stores across Shopee, Lazada, TikTok Shop, and Shopify with unified inventory sync.
High SKU Volume
45,000+ SKUs managed with variant support, bulk price updates, and automated stock alerts.
Sales Reporting & Analytics
Real-time revenue tracking (RM2.3M GMV/mo), profit margin analysis by channel, and performance insights.
End-to-End Automation
From product listing to order fulfillment — automated workflows that reduce manual work by 68%.
What Our Clients Say
"NeuralOps gave us enterprise-grade AI without the enterprise complexity. Our legal team uses it daily for contract review — data never leaves our VPS."
"The VPN-only architecture was the deciding factor. We process patient data — PDPA compliance isn't optional. AINNA delivered in 4 days."
"Cut our listing time from 3 hours to 45 minutes per store. The detached system runs on its own — our team focuses on strategy, not data entry."
Pitch 1: Scalable Framework for SME Transformation
Pitch 1 represents AINNA's expansion model — combining business experience, AI infrastructure, detached system development, and data-driven execution into a scalable framework for SME transformation.
Phase 1: Pilot
Single use case proof of concept with measurable ROI.
Phase 2: Rollout
Expand to multiple teams and integrate with core systems.
Phase 3: Scale
Organization-wide AI ecosystem with full automation.
Pitch 1 Framework
Scalable framework for SME transformation.
Phase 1: Pilot
Single use case proof of concept with measurable ROI.
Phase 2: Rollout
Expand to multiple teams and integrate with core systems.
Phase 3: Scale
Organization-wide AI ecosystem with full automation.
Pitch 1 Framework
Scalable framework for SME transformation.
Built for Real Industries
Secure AI infrastructure tailored to your industry's compliance and data sensitivity requirements.
E-Commerce
Product description generation, customer service agents, inventory forecasting, and automated listing management across multiple stores.
Legal Firms
Document analysis, contract review, case law research — confidential data stays behind your VPN. No third-party exposure.
Healthcare
Patient data processing, clinical report generation, medical record analysis — fully PDPA compliant, data never leaves your VPS.
Finance
Report generation, compliance monitoring, fraud detection, risk analysis — secure, auditable, with complete inference logging.
Government
Citizen services automation, document processing, policy analysis — sovereign AI infrastructure with Malaysia-based data control.
Manufacturing
IoT sensor monitoring, predictive maintenance, quality control automation, supply chain optimization with real-time agent execution.
From User to VPS to vLLM — Without Public Exposure
NeuralOps protects sovereignty by separating user access, agent execution, and LLM inference. Users reach the VPS/app layer. Only the designated VPS reaches vLLM through VPN.
User Stops at the VPS Layer
The user interacts with your app, dashboard, API, or agent running on VPS. The public side terminates here. The vLLM server is never exposed as a public destination.
Only Whitelisted VPS Can Enter VPN
The VPS uses WireGuard credentials and an approved IP route. If traffic does not originate from an authorized VPS tunnel, the vLLM layer rejects it.
vLLM Performs Inference Only
The LLM server receives a controlled prompt over VPN and returns output. It is not used as a storage layer, web app layer, or public integration point.
How NeuralOps Guarantees Data Sovereignty
Malaysia-Based Infrastructure
All servers and VPS instances are hosted within Malaysia's borders. Your data is subject to Malaysian law (PDPA 2010), not foreign jurisdictions like US Cloud Act or GDPR.
No Third-Party API Calls
Unlike solutions that proxy through OpenAI/Google/Grok APIs, NeuralOps makes zero external API calls. Your prompt never reaches a foreign server. Inference is 100% local to our infrastructure.
VPS-Level Data Isolation
Each customer's VPS is isolated at the cloud hypervisor level. No other tenant can access your files, processes, or memory. Your data stays inside your virtual boundary.
Zero-Knowledge Architecture
AINNA cannot see your data. We manage the infrastructure, not your content. The VPN tunnel and VPS encryption ensure your data is opaque to everyone except your authorized agents.
PDPA Compliance Built-In
Data sovereignty is the foundation of PDPA compliance. By keeping personal data within Malaysia and under your control, NeuralOps satisfies Section 129 (data transfer restrictions) automatically.
Audit Trail Without Data Exposure
All inference requests are logged for compliance — but only metadata (timestamp, token count, model used). The actual prompt content is never logged. You get auditability without exposure.
NeuralOps vs. The Alternatives
See how NeuralOps compares to OpenAI, Google Gemini, and self-hosted solutions across security, sovereignty, and control dimensions.
| Feature | NeuralOps | OpenAI API | Google Gemini | Self-Hosted |
|---|---|---|---|---|
| VPN-Only Access | ✅ Yes | ❌ No | ❌ No | ⚠️ DIY |
| No Public API Endpoint | ✅ Yes | ❌ Public | ❌ Public | ✅ Yes |
| Whitelisted VPS Only | ✅ Yes | ❌ API key only | ❌ API key only | ⚠️ Manual |
| WireGuard Encryption | ✅ Yes | ⚠️ HTTPS only | ⚠️ HTTPS only | ⚠️ Your setup |
| Data Stays in Malaysia | ✅ Yes | ❌ US servers | ❌ US/SG | ✅ Your choice |
| PDPA Compliance Ready | ✅ Built-in | ❌ GDPR only | ❌ GDPR only | ⚠️ DIY |
| No Training on Your Data | ✅ Guaranteed | ⚠️ Opt-out | ❌ May train | ✅ Yes |
| Audit Trail | ✅ Full | ⚠️ Limited | ⚠️ Limited | ⚠️ DIY |
| Zero Data Retention | ✅ Yes | ❌ 30 days | ❌ Varies | ✅ Yes |
| Centralized Management | ✅ AINNA manages | ❌ OpenAI controls | ❌ Google controls | ❌ You manage |
Defense in Depth — Multi-Layer Security Architecture
A layered security chart showing how the public internet is kept outside while only designated VPS nodes can reach the vLLM core through VPN.
Blocked
Whitelisted
Controlled execution
Frequently Asked Questions
Common questions about NeuralOps secure AI infrastructure.
Which LLM models are available on the central vLLM server?
We run 7 open-weight models optimized for different use cases: Llama 3.1 (8B/70B), Qwen 2.5 (7B/32B/72B), Nemotron 3 Ultra, and Mistral Nemo 12B. Model availability varies by plan — Starter gets 2 models, Business gets 5, Enterprise gets all 7. All models run locally on our GPU cluster; no external API calls.
What happens when I exceed my daily token limit?
Requests beyond your daily quota return a 429 response with a retry-after header. No overage charges — the limit resets at 00:00 UTC. Enterprise plans support custom rate limits. You can monitor usage via the VPS dashboard or request a limit increase through your account engineer.
What does "VPN-only access" actually mean for my team?
Your designated VPS receives a WireGuard config with a private IP (10.x.x.x). Only traffic originating from that VPS through the encrypted tunnel reaches the vLLM server. There is no public IP, no public DNS, no open ports on the LLM server. Your developers SSH into the VPS, run agents/apps there, and those apps call the private vLLM endpoint. The public internet cannot reach the inference layer at all.
How long does onboarding take and what's included?
Starter: 3–4 days. Business/Enterprise: 5–7 days. Includes: VPS provisioning (Malaysia DC), WireGuard tunnel setup, vLLM model allocation, DNS + SSL for your app subdomain, SSH keys, monitoring agent install, and a 1-hour handover call. Enterprise adds: dedicated account engineer, custom SLA review, and Hermes gateway integration if needed.
Is my data used to train or improve the models?
Absolutely not. Zero data retention on the vLLM server — prompts are processed in-memory and discarded immediately. No logging of prompt content (only metadata: timestamp, token count, model). Your VPS is isolated at hypervisor level; AINNA staff cannot access your files or memory. This is contractual and architectural.
What SLA do you offer and what's the uptime guarantee?
Enterprise: 99.9% uptime SLA with financial credits (pro-rata refund for downtime > 0.1%). Business: best-effort with priority support. Starter: community-tier monitoring. All tiers include: 24/7 infrastructure monitoring, automated failover for VPS layer, and snapshot-based recovery (RPO < 1 hour, RTO < 30 min).
Can I bring my own model or fine-tune on the central vLLM?
Enterprise plans support custom model deployment (GGUF / Safetensors) on dedicated GPU partitions — requires security review and 2-week lead time. Fine-tuning is not offered on the shared vLLM; we recommend running fine-tuning jobs on your VPS (we provide GPU-enabled VPS add-ons) and deploying the adapter/merged model to your dedicated partition.
How does NeuralOps compare to self-hosting vLLM on my own GPU server?
Self-hosting gives you full control but requires: GPU procurement (H100/A100 lead times), 24/7 ops expertise, VPN + hardening, model optimization, monitoring, and compliance auditing. NeuralOps offloads all infrastructure ops — you get a hardened, monitored, PDPA-compliant inference layer in days, not months. Cost comparison: RM3,500/mo (Enterprise) vs ~RM15K+/mo for equivalent self-hosted stack (GPU lease + colocation + engineering).
What support channels are available for each tier?
Starter: Email/ticket (48h response). Business: Slack/Teams + ticket (4h business hours). Enterprise: Dedicated Slack channel + phone + ticket (1h response) + quarterly architecture review. All tiers include access to runbooks, API docs, and the AINNA status page.
Where are the servers physically located?
All infrastructure — central vLLM GPU cluster and customer VPS instances — is hosted in Tier III data centers in Cyberjaya and Kuala Lumpur, Malaysia. Data never leaves Malaysian jurisdiction. This satisfies PDPA Section 129 (cross-border transfer restrictions) by default.
Start Your Pilot Project
Start with one workflow. Build one system. Scale the intelligence layer from there.
🔒 Enterprise-grade AI infrastructure with VPN-only security. No public API. No data leakage.
