The Oyvoda Engine

Five layers.
Not one model.

Most AI guest tools route every message into a single language model and hope for the best. Oyvoda runs five distinct, purpose-built layers before any response reaches a guest — each one engineered for the specific demands of short-term rental hospitality.

9
MCP Servers
<200ms
Avg Response
5
Architecture Layers
Zero
Ungovernered Actions

Every message takes a journey before it becomes a reply.

A guest asks about the pool heater at 11pm. Here's exactly what happens inside Oyvoda before they see a response.

SENSE
Emotional read
THINK
Knowledge retrieval
ACT
Governed execution
WATCH
Live observability
LEARN
Operator adaptation
01
Layer One · SENSE
We read emotion before we read intent.
Emotional Intelligence + Signal Detection
eq_mcp · signal_mcp

Before routing a guest message anywhere, Oyvoda runs it through an Emotional Intelligence layer that scores frustration, urgency, confusion, and sentiment. A guest asking "is the pool heated" at 2pm is different from a guest asking the same question at midnight with "we've been waiting 45 minutes."

Generic AI tools treat both messages identically. Oyvoda routes them differently — different tone, different urgency, different escalation threshold. Hospitality is fundamentally emotional. The engine reflects that.

Frustration Detection
Real-time scoring of guest emotional state. High frustration triggers escalation paths before the guest has to ask for a manager.
Intent Classification
14 distinct intent categories specific to STR: check-in, maintenance, local recommendations, pre-booking, emergency, and more.
Signal Aggregation
Market signals, property context, and guest history are bundled into a signal package before any response is drafted.
Urgency Scoring
Time of day, message cadence, and emotional markers combine into an urgency score that determines response speed and escalation chain.
02
Layer Two · THINK
Your property's knowledge. Not general knowledge.
RAG · Vector Store · Librarian Agent
knowledge_mcp · pms_mcp

Every Oyvoda operator gets a property-specific vector knowledge base — indexed from your house manual, PMS data, local guidebooks, and operator Q&A. When a guest asks a question, the engine retrieves the exact relevant chunks from your property's knowledge, not a generic training corpus.

This is the difference between "pool hours are typically 8am–10pm at most vacation rentals" and "your pool at 34 Seagrove Drive is heated to 84°F and available 24/7 — enjoy." Precision requires property-specific knowledge retrieval. Oyvoda builds and maintains that knowledge base automatically from your PMS data.

Property Vector Store
Each property gets its own indexed knowledge base. Retrieval is property-scoped — guests only get answers about their specific unit.
PMS-Native Sync
Live integration with Escapia, Guesty, and Track. Booking details, check-in instructions, and listing data sync automatically.
Librarian Agent
A dedicated agent curates and maintains the knowledge base — identifying gaps, resolving conflicts, and flagging outdated information.
KB Gap Detection
Every unanswered question surfaces in your operator dashboard. Approve an answer once and every guest gets it from that point forward.
03
Layer Three · ACT
Every action is governed before it executes.
Tool Execution · Governance · Sanitizer
governance_mcp · sanitizer_mcp

Oyvoda can take actions — draft replies, process late checkouts, log incidents, dispatch alerts, deliver gate codes. But every action passes through a hard-coded governance layer before execution. The architecture physically cannot take an action that hasn't been validated.

This isn't content filtering — it's structural safety. The 9 MCP servers that handle tool execution are registered at startup and cannot be modified at runtime. This eliminates prompt injection attacks, where a malicious guest message attempts to make the AI take unauthorized actions. The trust boundary is the architecture itself, not a rule set.

Hard-Coded Trust Registry
9 MCP servers registered at startup. No server can be added, removed, or modified at runtime. Eliminates prompt injection by design.
Pre-Booking Pipeline
Full AI-drafted pre-booking reply workflow. Drafts queue for operator approval before sending. Confidence scoring determines auto-send threshold.
Canary Rollout System
New AI behaviors roll out property by property, not portfolio-wide. One property advances to each phase before the next. Zero surprise deployments.
Escalation Handoff
When a situation exceeds the AI's scope, it escalates with full context — guest name, issue, conversation history, and urgency score — to the right person.
04
Layer Four · WATCH
You know before your guests notice.
SLO Monitoring · Real-Time Observability
watch_layer · slo_metrics

Oyvoda monitors every response for latency, accuracy, escalation rate, and knowledge retrieval quality — in real time, on every request. Service Level Objectives (SLOs) are defined per operator and tracked continuously. When the system degrades, the dashboard surfaces it before any guest has a bad experience.

This is the layer that makes reliability a measurable property of the system, not a marketing claim. Every operator sees their actual SLO metrics. Every knowledge gap is logged. Every escalation is tracked to resolution.

SLO Tracking
API latency (<1200ms), knowledge latency (<1500ms), error rate (<1%), and escalation backlog (<25) are monitored per operator, per property.
Operator Dashboard
Live view of all guest sessions, pre-booking queue, escalations, knowledge gaps, and performance metrics. One screen, everything that matters.
Alert Chain
Configurable escalation contacts per alert type per property. Primary contact → secondary contact → fallback, with timeout and acknowledgement tracking.
Incident Tracking
Property incidents — maintenance issues, guest complaints, compensation decisions — are logged, tracked, and reported. Full audit trail, zero spreadsheets.
05
Layer Five · LEARN
The longer you use it, the better it knows you.
Operator Learning · Preference Distillation
operator_learning

Every time an operator edits an AI draft — a word change, a tone adjustment, a policy clarification — Oyvoda analyzes the edit and updates that operator's preference profile. Over time, the system learns your specific voice, your specific policies, and your specific standards.

After 20 interactions, the AI writes drafts that need fewer edits. After 100, it writes like you. This is the layer that makes Oyvoda get permanently better for your specific operation — not just smarter in general, but smarter about you specifically. No other STR guest platform does this.

Edit Analysis
Every operator edit is classified by type: tone, policy, price signal, property fact. Each category trains a different aspect of the operator's AI profile.
Preference Distillation
Observed patterns distill into operator-specific instructions. "Always mention the beach walkover code in check-in confirmations" becomes a permanent rule.
Platform Intelligence
Anonymized, aggregated patterns across operators inform platform-wide improvements. Individual operator data is never shared.
Confidence Scoring
Each learned preference has a confidence score that grows with observation count. Low-confidence rules require approval. High-confidence rules apply automatically.
LLM Provider
Groq · Llama 3.1~200ms inference latency
Architecture
FastAPI + CeleryAsync Python, 3 worker queues
Database
Supabase Postgres21 tables, daily backups
Messaging
RCS-firstSMS fallback, no app download
PMS Integrations
Escapia · Guesty · TrackMore on request
Security
Hard-coded trust registryZero runtime MCP registration
Deployment
Railway · CloudflareAuto-deploy on every commit
Observability
Prometheus · GrafanaSLO monitoring, all services
Uptime
99.9% targetHealth checks every 15s

Built for hospitality.
Not retrofitted for it.

Horizontal AI tools handle support tickets for e-commerce and SaaS. Oyvoda is built from the ground up for the specific demands of short-term rental — where context is property-specific, timing is everything, and the guest experience is your reputation.

Oyvoda Generic AI Tools
Property-specific knowledge base Per-property vector store, auto-synced from PMS General knowledge only
Emotional intelligence layer Frustration, urgency, and sentiment scored before routing Topic-based routing only
Pre-booking AI Full Escapia/Vrbo inquiry pipeline with approval workflow Post-booking support only
Operator learning Learns your voice and policies from every edit Static behavior, no per-operator adaptation
Governed tool execution Hard-coded trust registry, zero runtime modification Content filtering only
STR-specific intents 14 STR-specific categories including pre-booking, maintenance, gate codes Generic support intents
PMS-native integration Escapia, Guesty, Track — live booking data, not manual entry Manual setup or CSV import
Canary rollout New behaviors test on one property before portfolio-wide All-or-nothing deployment

Built by operators.
For operators.

Oyvoda wasn't designed in a boardroom by engineers who've never managed a vacation rental. It was built in direct partnership with a 30A Florida STR operator managing 50 units — someone who has fielded the 2am WiFi calls, navigated the pet fee disputes, and spent hours answering the same pre-booking questions from Vrbo guests.

Every feature in the engine reflects a real operational pain. The pre-booking pipeline exists because inquiry response time directly affects Vrbo ranking. The emotional intelligence layer exists because a frustrated guest at checkout needs a different response than a curious guest at check-in. The operator learning layer exists because no two operators run their properties the same way.

This is what domain expertise looks like in software. Not a feature list. A system that reflects how the work actually gets done.

Real operator feedback shaped every layer of the architecture — from the escalation chain to the pre-booking confidence thresholds.
30A market expertise is embedded in the local knowledge layer — restaurant data, beach access, seasonal patterns, vendor relationships.
PMS integrations built to spec — Escapia, Guesty, and Track were chosen because they're what 30A operators actually use.
We built this because we needed it ourselves. Every operator managing more than 10 units is drowning in messages that don't need a human. Oyvoda handles the ones that don't, so you can focus on the ones that do.
O
Founding Operator
30A Florida · 50 Units · Escapia

How the engine works, plainly.

How does Oyvoda know about my specific property?
When you connect your PMS (Escapia, Guesty, or Track), Oyvoda syncs your property listings, booking data, and unit details automatically. You then complete a short operator questionnaire covering check-in rules, pet policies, local recommendations, and house rules. This data is indexed into a property-specific vector knowledge base — Oyvoda retrieves from your data, not a general training set.
What happens when the AI doesn't know the answer?
The AI doesn't guess. When retrieval confidence falls below threshold, the question surfaces in your operator dashboard as a Knowledge Gap. You approve an answer once, and every future guest asking the same question gets that answer automatically. The knowledge base grows with every interaction.
Can the AI make mistakes that embarrass my business?
The pre-booking pipeline operates in approval-required mode by default — every AI draft is reviewed by you before sending. As you build confidence in the system's accuracy for specific property and intent combinations, you can advance individual properties to auto-send mode. The canary rollout system means you never accidentally deploy a behavior change portfolio-wide.
How is this different from a chatbot?
A chatbot follows scripts. Oyvoda runs five distinct AI layers per message — emotional intelligence, property-specific knowledge retrieval, governed tool execution, real-time observability, and operator-specific learning. It handles the full pre-booking inquiry flow (before a guest has even booked), manages in-stay questions, detects and escalates emergencies, and improves with every operator interaction. No script could cover this range.
Is my property data private?
Yes. Each operator's data is isolated — scoped to their company_id at the database level. No operator's property knowledge, guest conversations, or booking data is accessible to other operators. Platform intelligence (anonymous aggregate patterns) is the only cross-operator data flow, and it contains no personally identifiable information.
What does "RCS-first" mean for my guests?
RCS (Rich Communication Services) is the upgraded messaging standard now supported on virtually all modern Android and iPhone devices. It delivers rich messages — read receipts, typing indicators, high-resolution images, interactive buttons — through the native Messages app. No app download, no signup, no friction for your guests. They just tap the link you send them. SMS is the fallback for older devices.

See the engine
working for your portfolio.

Connect your PMS and watch the first AI-drafted reply appear in your queue. Most operators are live within 48 hours.

Start Free Trial → ← Back to Platform