Mind Studios makes the project personal instead of just thinking of the income that comes from it. The team is very competent in what they do, always delivers on time, and, most important, is cost-efficient.
Unight
AI-Powered Venue Discovery Assistant
Industry:
Entertainment
Location:
China
Period of Collaboration:
October 2021 - March 2024
Tech Results:
Multi-source data integration
Custom ML tagging system
~3s query response time
LLM-powered intent recognition
Unight is Shanghai's leading nightlife discovery platform, serving 200,000+ users across hundreds of partner venues through a WeChat mini-program architecture.
Mind Studios was engaged to architect and develop the platform's AI-driven discovery layer.
What the client says about our first collaboration:
User experience as a strategic priority
Unight prioritizes user experience optimization above all strategic considerations. As Shanghai's leading nightlife discovery platform, their core mission centers on delivering accurate venue recommendations that align with individual preferences and situational context.
How venues store information
How users actually search
This represents a fundamental semantic gap. Users know how they want to feel, not which database category to filter by.
Market landscape
However, achieving this mission required solving a critical problem: conventional venue databases don't structure information the way users naturally search.
TripAdvisor and Dianping had reviews, but a generic category-based search
WeChat groups provided peer recommendations, but there was no searchable discovery tool
Local blogs offered insider insight, but weren't actionable or bookable
Our objective
We set out to build an AI system capable of interpreting emotional, contextual language and translating it into precise venue matches.
We architected a dual-layer AI system: one to build structured intelligence across every venue, and the other to translate user intent into precise and personalized recommendations.
Layer 1: Venue intelligence engine
Custom multi-dimensional classification framework processing unstructured data (reviews, descriptions, editorial content) into structured dimensional scores. Enables computational reasoning over experiential attributes.
Layer 2: Conversational orchestration system
LLM-based interface translating natural language queries into structured database operations through a six-step pipeline:
01.
Intent recognition
02.
Data parsing
03.
Vector search
04.
Re-ranking
05.
Guardrailed response generation
Outcome:
Personalized, context-aware recommendations delivered through a conversational interface, powered by a reliable data foundation, preventing AI hallucinations.
Technical implementation
AI layer #1:
Venue intelligence engine
This component is the foundation of our custom AI venue discovery platform development, not a simple tagging feature, but a proprietary dimensional classification ontology for nightlife venues.
Custom dimension framework
We developed proprietary dimensions that mirror how users naturally talk about nightlife:
Vibe
Overall atmosphere and character
Crowd composition
Demographics and social dynamics
Social density
Intimate vs. crowded environments
Conversation level
Background music vs. conversation-dominating
Music profile
Genre, volume, prominence
Energy curve
How the venue changes from early evening to late night
These dimensions capture subjective, contextual attributes that conventional categorical systems cannot represent.
Intelligence generation pipeline
Data sources
We aggregate content from multiple platforms to ensure comprehensive venue coverage:
TripAdvisor reviews and ratings
SmartShanghai editorial content
Internal user reviews
Community-generated commentary
Processing methodology
NLP models analyze unstructured text from these sources to:
01
Extract sentiment patterns
Identify positive/negative indicators within reviews
02
Detect contextual signals
Recognize nightlife-specific descriptors ("energetic," "intimate," "sophisticated")
03
Score dimensional attributes
Assign quantitative values to each proprietary dimension
Output:
Each venue becomes a rich data object with dimensional scores, not just a name and category label.
Technical implementation
AI layer #2:
Conversational AI concierge system
With structured venue intelligence established, we developed a natural language interface that translates user queries into precise recommendations. The result is an AI venue search platform development approach that handles nuanced and multi-intent queries at production scale.
The system executes six operations in sequence
01
Intent recognition
02
Context extraction
03
Dimension mapping
04
Semantic retrieval
05
LLM-based re-ranking
06
Guardrailed response generation
Architectural principles
The conversational layer was designed with:
-
Separation of intelligence and conversation layers
-
Deterministic data foundation
-
Explainable ranking logic
-
Latency optimization for mini-program constraints
-
Scalable embedding infrastructure (pgvector)
-
Modular AI services enabling future expansion
Semantic search architecture
Technology:
PostgreSQL database with pgvector extension for vector similarity search
How it works:
Traditional keyword search matches literal text: searching "bar" returns venues labeled "bar." Our semantic search operates differently:
-
Each venue's dimensional scores convert into a mathematical vector (array of numbers)
-
User queries also convert to vectors based on detected preferences
-
System finds venues with similar vector patterns, not keyword matches
Example:
Query:
Somewhere relaxed for conversation
System matches:
Venues with high conversation-level scores, moderate energy curves, and appropriate social density, regardless of their category labels.
Why this architecture matters
Computational reasoning
The system processes queries mathematically across dimensions, enabling complex matching that keyword search cannot achieve.
Explainable recommendations
Each suggestion can be explained through dimensional alignment: "This venue matches your preference for moderate energy and conversation-friendly atmosphere."
Controlled outputs
Ranking logic operates on structured dimensional data, ensuring all recommendations derive from verified venue attributes.
Scalable intelligence
New venues automatically process through the same pipeline, maintaining consistent classification quality as the database grows.
Technical execution
User input:
Celebrating with coworkers, good music, but we can talk, not too expensive.
What the system executes:
01
Intent classification
~150ms
System identifies:
- Venue discovery intent
- Group context
- Budget constraint
- Social energy parameters
Parallel processing branches activate if the query contains multiple intents.
02
Query decomposition
~200ms
Extraction of structured components:
-
- Social atmosphere: professional, group-friendly
-
- Music: present but not dominant
- Conversation: enabled
- Price sensitivity: mid-range
- Energy level: moderate
- Exclusions: high-intensity nightclub environment
03
Vector search
~1.2s
Query converts to embedding vectors. Search executes against pre-indexed venue embeddings via pgvector. Tag filters refine the result set.
04
LLM re-ranking
~150ms
Candidate venues evaluated for contextual alignment using:
- Tag proximity scoring
- Internal quality metrics
- Review sentiment alignment
- Historical engagement signals
05
Guardrailed response generation
~650ms
LLM constructs explanations using exclusively verified database entities. No external inference. Strict grounding to structured venue data prevents hallucination.
Total execution:
~3 seconds
Want recommendations that feel personal, not algorithmic?
We design conversational AI that guides users naturally from question to decision.
The AI implementation transformed Unight's fundamental platform architecture.
Before
Transactional ticketing utility
After
AI-driven discovery ecosystem
Impact areas
User behavior
Transactional sessions → Exploratory browsing
Users engage for discovery independent of immediate booking intent.
Decision confidence
Multi-platform comparison → Single conversational interface
Natural language queries eliminate research fragmentation.
Monetization
Low engagement → Increased venue exposure and partnership value
Extended sessions strengthen advertising effectiveness.
Data intelligence
Anonymous visits → Structured behavioral signals
Every interaction refines recommendation accuracy.
We delivered a production-grade AI system serving 200,000+ users with sub-3-second natural language venue discovery. The modular architecture provides Unight with a scalable foundation supporting future user base expansion and venue network growth.
Want an AI solution that matches your users' natural behavior?
Book a free consultation to explore how conversational AI can transform your platform.
Rebind: AI-Powered e-Reading Platform
Next case
