The AIgent v 0.1
Using AI to connect writers & representation
Challenge
Today, literary publishing is a $20 billion industry, with a serious bottleneck problem. Successful agents receive on the order of hundreds, if not thousands, of query letters daily. Before they can even sit down to the important task of reviewing a manuscript, hours of the day are devoted to scrolling through endless submission emails, and sorting the wheat from the chaff.
App Solution
The AIgent is an app designed to solve this problem, by helping literary agents rapidly identify pitches that are a good match with their current portfolio. It leverages state-of-the-art techniques from ML and NLP to pinpoint the genre a pitch falls into, and automatically contextualizes it in terms of similar titles. In a centralized agency workflow, these techniques could be scaled to route submissions directly to agents that best match the query letter, saving agents (and writers!) time, money, and frustration.
Repo Overview
Build Workflow
The AIgent was built in a 3-week period as part of an Insight Data Science fellowship.
- Scraped the metadata tags and synopses of >100k GoodReads titles using BeautifulSoup.
- Leveraged pre-tained embeddings from the state-of-the-art BERT language model.
- Built out a text classifier and transfer learning pipeline with TensorFlow, sklearn, and Pytorch.
- Developed a robust classifier, achieving > 85% accuracy and > .9 F-scores, across 20 genres.
- At inference, identified similar titles on the basis of embedding cosine similarity.
- Deployed model as a Flask web app, hosted on AWS.