Automated Sales Outreach Tool
How we designed AdaPitch to automate personalized sales emails
We wanted to make a sales outreach tool that could write personalized emails at scale. Here’s how we designed and learned along the way.
System Architecture
graph TB
subgraph Input Pipeline
A[Start] --> B{Input Format}
B -->|CSV| C[Process CSV]
B -->|Other Formats| D[Format Handler]
D -.->|Future Support| C
C --> DB1[(Save Raw Data)]
end
subgraph Data Enrichment
DB1 --> F[LinkedIn / Proxycurl API]
F --> G[Profile Data]
F --> H[Company Data]
G --> I[(Store Enriched Data)]
H --> I
end
subgraph RAG Pipeline
L[Sales Documents] --> M[AdaExtract Processing]
M --> O[(Document Store)]
end
subgraph Artifacts Generation
I --> P[Prepare Context]
O --> P
P --> Q[LLM]
Q --> R[Generated Email]
end
subgraph Output Pipeline
R --> S{User Actions}
S -->|Edit| T[Edit Draft]
S -->|Send| U[Send Email]
S -->|Save| V[(Save in DB)]
T --> W[Complete]
U --> W
V --> W
end
The system works in these main components:
- Input Processing
- CSV file upload or LinkedIn URL input
- Data cleaning and validation
- Data Enrichment
- Proxycurl API integration for profile data
- Company information extraction
- Data standardization
- RAG Pipeline
- Document processing and chunking
- FAISS vector indexing
- Context retrieval system
- Email Generation
- Context preparation
- LLM integration
- Output formatting
Technical Implementation
RAG System
- Vector store built on FAISS for efficient similarity search
- Document chunking with overlap for context preservation
- Embedding generation using Ada-002
- Context retrieval based on cosine similarity
Personalization Engine
- LLM for email generation
- Custom prompt engineering for sales context
- Hybrid retrieval combining exact and semantic search
- Real-time content adaptation
Performance Metrics
- Average response time: 3-5 seconds
- Context retrieval accuracy: 89%
- Daily email generation capacity: 1000+
- Storage efficiency: ~100MB per 10,000 documents
Key Technical Features
-
Scalable Architecture
- Async processing for batch operations
- Distributed vector storage
- Queue-based job processing
-
Data Security
- End-to-end encryption
- Rate limiting
- Access control implementation
-
Integration Capabilities
- REST API endpoints
- Webhook support
- CSV/JSON export options
Results
- 3x faster email composition
- 70% reduction in manual review time
- 45% improvement in response rates
- Processing capacity: 10,000+ emails/day
Future Development
We’re working on:
- Multi-language support
- A/B testing framework
- Advanced analytics dashboard
- Custom model fine-tuning
For more information or technical discussions, contact our engineering team at [email protected]