Prosus developed Plus One, an internal LLM platform accessible via Slack, to help companies across their group explore and implement AI capabilities. The platform serves thousands of users, handling over half a million queries across various use cases from software development to business tasks. Through careful monitoring and optimization, they reduced hallucination rates to below 2% and significantly lowered operational costs while enabling both technical and non-technical users to leverage AI capabilities effectively.
# Plus One: Prosus's Journey in Enterprise LLM Deployment
## Overview
Prosus developed Plus One as an internal LLM platform to serve their diverse portfolio of companies reaching approximately two billion users. The platform acts as an AI team member, providing various capabilities through a Slack interface, making AI tools accessible across their organization.
## Platform Capabilities and Architecture
### Core Features
- Multi-model integration supporting various tasks:
### Technical Implementation
- Slack-based interface for easy access
- Integration with internal knowledge bases (Confluence)
- Support for multiple underlying LLM models
- RAG implementation for improved accuracy
- Custom feedback system using emoji reactions
## Usage Patterns and Adoption
### User Demographics
- Thousands of active users
- Over half a million queries processed
- 50/50 split between engineering and non-engineering users
### Common Use Cases
- Software development tasks
- Product management
- General writing and communication
- Technical documentation
- Financial analysis
- Marketing tasks
## Performance Optimization
### Hallucination Reduction
- Initial hallucination rate: ~10%
- Current rate: Below 2%
- Improvement strategies:
### RAG Implementation Challenges
- Context-specific optimization required
- Parameter sensitivity:
## Cost Optimization Strategies
### Token Economy
- Continuous cost monitoring per task
- Model selection based on task requirements
- Language-specific considerations:
### Cost Reduction Methods
- Model size optimization
- Efficient prompting
- Context length management
- Strategic model selection
- Fine-tuning when necessary
## Production Deployment Learnings
### Key Insights
- Value discovery through experimentation
- Balance between innovation and practical applications
- Scale considerations for production deployment
- Cost-effectiveness at scale
### Implementation Strategy
- Start with prompt engineering before fine-tuning
- Focus on existing problem solutions
- Careful consideration of model selection
- Progressive scaling approach
## Agent Implementation
### Current State
- Early experimental phase
- Success in specific use cases
- Challenges:
### Quality Control
- Human-in-the-loop verification
- Team-based quality control through Slack
- Domain expert review for specific applications
- Continuous monitoring and feedback collection
## Business Impact
### Productivity Improvements
- Faster task completion
- Skill democratization
- Reduced bottlenecks
- Enhanced communication
### Use Case Examples
- K-12 education platforms
- Food delivery services
- Technical documentation
- Code assistance
- Content generation
## Risk Management
### Safety Measures
- Feedback collection system
- Human oversight
- Team-based verification
- Domain expert review
- Continuous monitoring
### Challenges Addressed
- Information accuracy
- Model reliability
- Context-specific requirements
- Scale considerations
- Cost management
## Future Considerations
### Areas for Improvement
- Agent reliability
- Cost optimization
- Multi-language support
- Context handling
- Information freshness
### Ongoing Development
- Model selection optimization
- RAG implementation refinement
- Cost reduction strategies
Start your new ML Project today with ZenML Pro
Join 1,000s of members already deploying models with ZenML.