3 High-level overview

3.1 Core steps of a peaks and project

Peaks and pits have gone through many iterations throughout the past year and a half. Currently, the general workflow is

1. Extract brand/product mentions from Sprinklr (the start of any project)
2. Classify a sample of posts (of each sentiment) using GPT-3.5 OR one of the latest SetFit model that has been developed (e.g. the one from the previous project) to quickly identify peaks and pits
3. Human review to select exemplar peaks and pits from these ‘crudely identified posts’
4. Fine-tune the SetFit model using selected exemplar posts (from current project and previous projects)
5. Run inference of this fine-tuned model over all of the project specific data
6. Use GPT-3.5 for an extra layer of classification on identified peaks and pits
7. Further understanding the identified peak and pit moments
   • 7.a. Conduct topic modelling via BERTopic over peaks and pits separately to identify high level topics for each brand x peak/pit
   • 7.b. Utilize GPT-3.5 for multilabel classification of Brand Love Emotion States in peak and pit posts

Figure 1.1: Schematic workflow from Project 706