Kevin Pillsbury

What is this: Write or paste a book or story synopsis into the input field, press the "Rate" button, and a trained neural network will rate your synopsis! How it works: You input a synopsis, my website takes your synopsis and makes an API request to a Google embedding model. Google's embedding model takes your synopsis, transforms it into a vector embedding, then returns that embedding to my website. My website takes this embedding, and runs it through a neural network I built and trained which returns a quality rating of the embedded synopsis. How I built it: Rating a story's synopsis involves two distinct parts. Part A converts the synopsis into an embedding (a numerical vector), while Part B converts that embedding into a rating. Part A requires a large natural language AI model, for which I lack the training data and computing power to build from scratch. Part B is a "head"—a small, task-specific neural network added to a base model—which I do have the capacity to build. I could have downloaded a pre-trained model for Part A, but it would be too large to run on this website, it would require a second server and an API to connect to it. Instead, I am using Google's state-of-the-art embedding model via their API. For Part B, I found a dataset of ~100k book descriptions and their corresponding Goodreads ratings. Using TensorFlow, I trained a small neural network (the head) to take a synopsis embedding and generate a rating. During training, I used the Google API to embed the descriptions to ensure consistency with how I process user synopses. My neural network head is small enough to run locally on this website, eliminating the need for a second server.