AI systems.
Software and hardware, delivered together.
AI hospitality booking, AI customer service, AI ERP. MAQ handles the design, development, and deployment. When an AI server is needed, MAQ delivers that too.
Three systems. All live in production.
Connect your booking system, and the AI can check real availability.
AI Hospitality Booking
Connects to your booking system and rate calendar. Checks availability, calculates rates, and generates booking links.
- Real-time availability, including weekday/weekend rates, extra guests, pets, and breakfast rules
- Booking links pre-filled with dates and guest count, discount codes supported
- Public holidays calculated automatically, never guessed by the model
- Price negotiations and complaints route to a human, with instant owner notification
- Multiple LINE Official Accounts sharing one rule set
AI Customer Service
The model runs on your own hardware. It answers what it can, and hands off to your team when it can't.
- Conversations stay on your own server, never routed through a third-party service
- Tool calls query your internal database and knowledge base directly
- Human handoff: push notifications, one-tap switch-over, nothing missed
- Deploys on both a website widget and a LINE Official Account
- Tone, permissions, and reply rules are all configurable in the admin panel
AI ERP
Query your inventory, purchasing, and orders in plain language. Data stays in its original system.
- Natural-language queries for customers, orders, inventory, and purchasing
- Connects to your existing tables, no separate data warehouse needed
- Knowledge-base retrieval, so new hires can just ask
- Tiered permissions keep sensitive data out of external channels
- Can be deployed entirely on your own server
Kei Café B&B.
Eight properties, one system.
Late-night and holiday inquiries used to wait until the next day. Now the AI checks availability, quotes rates, and sends booking links directly. Price negotiations and complaints get routed to a human.
The system runs on their own server. Conversations and guest data stay on their local network.
What it does. More than just answering.
It checks, it calculates, it sends links. And it knows when to hand off to a person.
It knows your real availability.
Connected to your booking system for real availability. Weekday/weekend rates, extra guests, pets, and breakfast are all calculated from your rate calendar
Links that open straight through.
Dates and guest count are pre-filled — click through straight to the quote page. Discount codes apply automatically
It understands holidays.
Dragon Boat Festival, Mid-Autumn Festival — converted to the correct dates automatically. Calculated by code, never guessed by the model
It never negotiates on its own.
If a guest mentions budget or pushes back on price, it asks questions first and notifies the owner. The AI never lowers the price itself
Safety comes before anything else.
When damage is reported, it first checks whether anyone was hurt. Compensation is then handed off to a person
A person can take over anytime.
Ask for a human and a push notification goes out instantly — the AI goes quiet. One tap in the admin panel switches back, nothing gets missed
Eight accounts, one brain.
Each property has its own LINE Official Account and website widget, all backed by the same rules and data
The restaurant, too.
Checks breakfast and brunch hours, with signature dishes and how to reserve
Performance. Measured on live hardware.
Every number below was measured on this production server. Model: gpt-oss-120B.
Response time. Concurrent capacity.
Each turn ingests roughly 10–12K tokens of context and conversation. The system decides on its own whether a database lookup is needed.
| Conversation type | Measured | Notes |
|---|---|---|
| Facility questions | 2–4 sec | Answered directly, no database lookup |
| Availability & rate checks | 6–10 sec | Calls a tool to check live availability |
| Cross-property comparison, full buyout | 10–14 sec | Scans multiple properties at once |
| Configuration | Concurrent conversations | Experience |
|---|---|---|
| Standard | 4 guests | Each keeps normal speed |
| Current 8-slot setup | 8 guests | VRAM headroom to spare |
| Theoretical max | ~600–900 sessions/hr | 3–5 turns per session |
// Test environment: Ollama + gpt-oss-120B, single NVIDIA RTX PRO 6000 96GB, ~3ms local network latency.
Proof.
We use it ourselves first.
Every architecture we recommend to clients has already run on our own production systems.
- Our own ERP, built by us.
- Inventory, purchasing, orders, e-invoicing, cost reports, and staff scheduling — all built in-house and running in production for over fifteen years. When we talk about enterprise systems, MAQ is a user, not just a vendor.
- Our own AI Consultation, running on our own hardware.
- MAQ's AI Consultation runs on an on-premises AMD Radeon AI PRO R9700 host, using the open-weight Gemma 4 model. The same architecture, producing real, measurable performance.
- Software and hardware, one team.
- System development, interface design, model deployment, and server builds all happen at MAQ. When something goes wrong, there's no relay between a software vendor and a hardware vendor.
For the hospitality industry. Get in touch.
Questions cluster around late nights and long holidays — exactly when staffing is hardest to cover. The same system adapts to your business type.
Start with what you have.
Where your booking, POS, and customer data already live, and which questions get asked every day. No need to replace existing systems.
Decide where it runs.
Choose cloud API or your own hardware based on data sensitivity and usage. Most start on the cloud and move to an on-premises AI server as usage grows.
Connect systems, build the interface.
Integrate your booking system and rate rules, set up LINE and website entry points, and define when to hand off to a human.
Launch with one.
Start with a single brand or time slot. Once reply quality and conversion are confirmed, expand to every location.
The AI server.
Delivered alongside it, when you need one.
No more worrying about cloud token costs, or policies that require data to stay in-house — the entire system can run on your own server.
A single 96GB GPU fully loads a 120-billion-parameter model — no splitting, no reduced precision. One machine has enough compute to serve multiple brands at once.
RTX PRO 6000
96GB of VRAM, loads a 120B model on a single card
Threadripper PRO
32-core server platform with IPMI remote management
DDR5 ECC
Server-grade error-correcting memory for stable long-running workloads
Ubuntu + Ollama
OpenAI-compatible API, model stays resident in VRAM
Get started. Let's talk about your workflow first.
The first step isn't picking a model — it's deciding which tasks the system should handle, and which data can't leave your internal network. If the numbers don't work out, MAQ will tell you plainly.