Hotels run on constant communication: guest calls, housekeeping requests, late checkout questions, booking changes, complaints, confirmations, and staff coordination. The challenge is not simply answering a guest. The challenge is turning a natural-language request into a safe, structured, and traceable hotel operation.
Back to portfolio
Voice AI / Agentic Automation / Hospitality TechCompany product / Current work
Operaios: Voice AI for Agentic Hotel Automation
Operaios is a company-built hotel automation platform designed to automate guest calls and hotel operations through voice-first AI workflows. I own the backend, voice AI layer, agent workflows, integration logic, and AI engine responsible for turning natural-language guest requests into structured hotel operations.
Details are sanitized because this is current company work.
Context
Hotel operations are communication-heavy systems.
Hotels run on constant communication: guest calls, housekeeping requests, late checkout questions, booking changes, complaints, confirmations, and staff coordination. The challenge is not simply answering a guest. The challenge is turning a natural-language request into a safe, structured, and traceable hotel operation.
Problem
Most hotel workflows break between conversation and execution.
Most hotel workflows are fragmented. A single guest request can require understanding the request, checking context, applying policy, triggering an internal workflow, notifying staff, confirming with the guest, and logging the action. Basic chatbots can answer questions, but they do not reliably execute operational workflows.
Most hotel workflows are fragmented. A single guest request can require understanding the request, checking context, applying policy, triggering an internal workflow, notifying staff, confirming with the guest, and logging the action. Basic chatbots can answer questions, but they do not reliably execute operational workflows.
My ownership
The language here matters because it should reflect real scope.
I own the backend, voice AI, and agentic AI engine for Operaios. My work focuses on the system that listens to guest requests, understands intent, plans the workflow, calls backend tools, coordinates hotel operations, and records the result.
I own the backend, voice AI, and agentic AI engine for Operaios. My work focuses on the system that listens to guest requests, understands intent, plans the workflow, calls backend tools, coordinates hotel operations, and records the result.
System architecture
From guest voice request to operational action and audit log.
The system is structured around understanding, planning, execution, confirmation, and traceability.
System architecture
Voice request -> planning -> action -> audit trail
01
Guest voice request
02
Speech understanding
03
Intent detection
04
Guest/context lookup
05
Agent workflow planner
06
Backend tool calls
07
Hotel operation execution
08
Confirmation response
09
Event/audit log
10
Human escalation when needed
Natural language in. Controlled operations out.
Example workflow
A single guest request can trigger multiple coordinated backend actions.
This is where voice AI stops being a chat demo and starts behaving like an operational system.
Guest says
“Can I get late checkout and extra towels?”
Step 1
Detects two intents: late checkout and housekeeping request
Step 2
Checks guest/context information
Step 3
Validates whether late checkout can be handled automatically or needs staff approval
Step 4
Creates or routes housekeeping task for extra towels
Step 5
Confirms the result to the guest
Step 6
Logs the full interaction for staff visibility
Engineering challenges
The hard parts are reliability and decision quality under real operating conditions.
These are the constraints that shape architecture choices when the system is responsible for real guest-facing actions.
Low-latency voice response
Interruptions and unclear speech
Intent detection accuracy
Agent state management
Tool-call safety
Avoiding wrong reservation or guest-service actions
Human escalation thresholds
Backend reliability
Event logging and auditability
Testing voice-agent workflows
Senior engineering decisions
Production AI systems need guardrails, not just prompts.
These are the design principles that make agentic systems safer to run in the wild.
Tool boundaries
Agents should only execute approved actions through controlled backend tools.
Human escalation
Sensitive or uncertain requests should be escalated instead of guessed.
State management
Voice workflows need memory of what was said, confirmed, denied, and completed.
Audit trail
Every automated action should be visible and explainable later.
Latency
Voice AI must feel realtime; slow responses break user trust.
Evaluation
Agent workflows should be tested against repeatable scenarios, not judged by one demo.
What I learned
The strongest lesson is that safe automation beats impressive improvisation.
Voice-first agentic systems succeed when they reduce operational noise without hiding decisions.
Voice AI is harder than chat because latency, interruptions, and recovery matter.
Agentic systems need strict backend tool boundaries.
Hotel automation requires traceability because actions affect real guest experiences.
Human escalation is not a weakness; it is part of safe production AI.
The best AI systems reduce operational noise without hiding important decisions.