pranav.
Back to articles
Machine LearningNLPStartups

Building ML Models for Sales Performance: Lessons from RillaVoice

October 12, 2024·10 min

Building ML Models for Sales Performance: Lessons from RillaVoice

During my time at RillaVoice (a Bessemer-backed startup), I built an unsupervised topic algorithm for analyzing sales conversations. Here's how I achieved a 72% improvement in accuracy.

The Challenge

Sales conversations are messy:

  • Multiple languages and accents
  • Small, imbalanced datasets
  • Domain-specific terminology
  • Real-time processing requirements

    • Traditional topic modeling (LDA, LSA) failed because they rely on large, clean datasets.

    The Solution Stack

    I combined multiple NLP techniques:

    1. BERT Embeddings

  • Contextual word representations
  • Better handling of sales jargon
  • Transfer learning from large corpora

    • 2. Dialogue Act Tagging

  • Distinguish questions from statements
  • Identify objection handling
  • Track conversation flow

    • 3. Part-of-Speech Filtering

  • Focus on nouns and verbs
  • Remove filler words dynamically
  • Preserve domain terms

    • 4. Topic Coherence Measures

  • C_v score for evaluation
  • Automated hyperparameter tuning
  • A/B testing different approaches

    • Results & Impact

    The improved model:

  • Processed Pepsi and Verizon sales calls
  • Generated custom performance suggestions
  • Ranked salesperson performance automatically
  • Scaled to multiple languages

    • Key Takeaways

  • **Domain expertise matters**: Understanding sales conversations was more valuable than ML theory
  • **Start simple**: Baseline models help you understand the problem
  • **Iterate quickly**: Small improvements compound
  • **Business impact > Academic metrics**: Focus on what drives value

    • This project taught me that the best technical solutions come from deep understanding of the business problem.

    ← View All Articles