# QuantumBlack's Dual LLM Applications Case Study

This case study covers two significant LLM applications developed by QuantumBlack: molecular discovery for pharmaceutical research and call center analytics for banking clients. Both applications demonstrate sophisticated LLMOps practices while addressing different industry challenges.

‍

## Molecular Discovery System

‍

### System Overview

- Developed for pharmaceutical and biotech research applications

- Combines chemical language models with RAG capabilities

- Processes scientific literature and molecular databases

- Uses vector databases for efficient information retrieval

- Supports multi-modal processing including text and chemical structures

### Technical Implementation

- Chemical Language Models:

- RAG Implementation:

### Production Considerations

- Built with domain-specific requirements in mind

- Supports processing of complex chemical notations

- Handles multiple data representations (SMILES, graphs, etc.)

- Incorporates chemical validation rules

- Scales to process large molecular databases

## Call Center Analytics System

### Architecture Overview

- Batch processing pipeline for historical audio files

- Kubernetes-based deployment

- Hybrid cloud/on-premises architecture

- Four main pipeline components:

### Technical Components

### Diarization Implementation

- Initially used PyAnnote for speaker detection

- Optimized using domain knowledge of call center audio format

- Switched to Silero VAD for efficiency

- Achieved 60x speedup through:

### Transcription Service

- Uses OpenAI Whisper model

- Implemented custom batching

- Distributed processing using Horovod

- Optimizations:

### LLM Analysis

- Uses Mistral 7B model (4-bit quantized)

- Multiple inference passes for consistency

- Structured output generation

- Custom prompt engineering

- Polling mechanism for numerical assessments

### MLOps Infrastructure

- MLRun Framework Usage:

- Production Considerations:

### Performance Optimizations

- Resource Utilization:

- Scalability Features:

### Output and Integration

- Structured Data Generation:

- System Integration:

## Production Deployment Considerations

### Security and Compliance

- PII detection and anonymization

- Regulatory compliance support

- On-premises deployment options

- Data privacy controls

### Scalability and Performance

- Optimized resource utilization
- Parallel processing capabilities
- Efficient data handling
- GPU resource management

### Monitoring and Maintenance

- Pipeline status tracking
- Performance metrics
- Error handling
- Resource utilization monitoring

### Future Extensibility

- Support for new models
- Additional language support
- Enhanced analytics capabilities
- Integration with other systems