← All projects

Sales Lead Proposal System

A lead qualification and proposal engine that predicts purchase intent using semantic similarity rating rather than asking an LLM to score leads directly.

2025LLMsEmbeddingsSemantic SimilarityPurchase IntentPython
Status
In production
Role
Implementation

Ask an LLM to rate a sales lead from 1 to 5 and it will happily give you a number. The problem is the distribution: models cluster on a couple of values, avoid the extremes, and shift their scale depending on how you phrase the question. The scores look reasonable one at a time and fall apart in aggregate.

This system takes a different route, implementing Semantic Similarity Rating (SSR) — a method introduced by Maier et al. in LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings.

Rather than eliciting a number, SSR has the model respond in natural language, then maps that response onto a rating distribution by comparing its embedding against a set of reference statements anchored to each point on the scale. The rating is recovered from semantic position, not asserted by the model.

Two properties made it worth building on:

Applied to sales, that turns qualification from an opaque score into something a human can audit and act on, and feeds the proposal generation that follows.

Credit: the SSR method is the work of Maier, Aslak, Fiaschi, Rismal, Fletcher, Luhmann, Dow, Pappas, and Wiecki. My work was implementing and adapting it for lead qualification and proposal generation.

innoscripta SE · 2025