Salesfloor’s AI associate uses proprietary Shopping Intelligence to enable real-life autonomous conversations with shoppers engineered for a personalized brand experience and conversion.

Role I played:

I owned end-to-end development of the conversational AI engine. Here’s how I drove the project technically and strategically:

1. Core System Design

Problem:

Traditional retail chatbots failed because:

• They couldn’t handle nuanced shopping dialogues
• Lacked real-time personalization
• Felt robotic vs. human sales associates

My Solutions:

• Architected a hybrid conversational AI system combining:
  • Rule-based workflows for critical retail scenarios (e.g., returns, sizing)
  • Generative AI (fine-tuned GPT-3.5) for natural responses
  • Reinforcement Learning to optimize for conversion
• Designed the Shopper Context Graph:
  • Real-time integration of:
    • Browse behavior
    • Purchase history
    • CRM data (loyalty tier, past returns)

Key Technical Contributions

A. Intent Classification

• Trained a DistilBERT model on:
  • 500K+ retail chat transcripts
  • 38 shopping-specific intents (e.g., “fit advice”, “price negotiation”)
• Achieved 98% accuracy by adding retail-specific embeddings

B. Conversational Memory

• Implemented a hierarchical attention mechanism:
  • Short-term memory (last 3 messages)
  • Long-term memory (shopper profile)
  • Product catalog attention

C. Conversion Optimization

• Built a custom RL environment where the AI learns from:
  • Positive signals (add-to-cart, checkout)
  • Negative signals (message ignores)
• Used PPO (Proximal Policy Optimization) for stable training
• Built on Neo4j for rapid relationship queries