Overview

This was a passion project I took on entirely on my own. I identified a major pain point, came up with a solution, and built it myself.

The project was a custom GPT designed to let users search and query Fox Nation's CMS data using natural language. It solved for gaps in how content editors searched for and discovered content, and gave them easy access to engagement data that was previously hard to get.

The Problem

There were two major pain points with Fox's internal CMS, called DPP:

First, searching for content was nearly impossible unless you knew the exact title. The CMS search didn't extend beyond title metadata, so searching by genre tags, description text, or anything else just didn't work. Content editors had a really hard time with this when putting together editorial collections.

Second, pulling engagement data from our systems required going through the Business Intelligence team, who had an obscure method for extracting raw CMS data (collection impressions, show and episode views, etc.). The data came back messy and required formatting knowledge to understand. This made it very challenging for content editors to make data-driven decisions when organizing content.

How I Got Started

Fox has a voluntary mentorship program that pairs people below Director-level with senior leaders outside their part of the org. I got matched with David, an SVP who led the team building FoxGPT, an internal version of ChatGPT that let Fox employees use an LLM with proprietary information without that data going back to OpenAI.

David and I talked a lot about my career goals and my interest in product management. At one point, he mentioned his team was working on letting employees build their own custom GPTs. When they started beta testing, he asked if I wanted to be included. Of course I said yes!

From there, I got access, spent some time brainstorming pain points I could address, and landed on this idea. Then I just went and built it.

How It Works

The GPT had CMS data fed into it as a knowledge base. This included all show, episode, and movie metadata: titles, descriptions, genre tags, category tags, and more. It also had engagement data under the hood, so it could answer questions about metrics for shows, episodes, movies, and content collections.

Users could ask questions in natural language, and the GPT would parse through the data and return a response. For example, I could ask "Give me a list of every piece of content related to George Washington," and it would quickly return a full list after searching the entire content library. Previously, this was a very manual effort. Content editors had to comb through the whole library or rely on their memory of the available content.

Building the Knowledge Base

I started by having our BI team pull raw CMS data in Excel format. This let me see all the field names under the hood. From there, I gave the GPT context around what each field name represented, so it had a clear understanding of what each data field meant.

I tested results against the Excel data pulls to make sure I felt confident in the GPT's output before moving on to the API integration.

API Integration

I used Custom Actions to integrate the DPP API endpoint, which let the GPT query live CMS data rather than relying on static exports. I followed documentation to get most of the way there, and enlisted help from a Data Scientist on David's team (someone I had previously worked with on the ML recommendations project) to fill in my knowledge gaps. I had a lot of fun working through this process and learning a bit about API integrations!

The Before & After

Before - Virtually no ability to search for content easily. Lots of manual, time-consuming digging through the CMS. Needed to lean on another team to pull engagement data, which came back raw and messy.

After - Users could type a simple prompt to pull entire lists of content seamlessly, taking a fraction of the time and effort. They could also ask the GPT questions about engagement data and get clean, easy-to-read answers very quickly.

Status

Unfortunately, this remained a prototype. I left Fox for my current role at Grayscale before I was able to ship it to production. I was the only user of the prototype, so I never got external feedback on it.

Additional Thoughts

I'm proud of the fact that I went out and did this entirely on my own, purely to solve a major pain point my team dealt with regularly. At its core, product management is about creative problem solving. This was a bit of a passion project, but it helped show me that my natural desire to solve problems and optimize workflows meant I was on the right path pursuing a career in product management.

This was my first solo foray into building an AI product, and I had a blast with it! Looking back, the main thing I took away is that I was able to make this happen because I had a self-starter attitude and just went for it.