This case study is based on a real project but fictionalized for client's confidentiality. We used public available information to demonstrate how a project with Gravitate AI works.
Our hypothetical client (Client) provides comprehensive software solutions to large corporations. They started to realize AI adds competitive advantages. As a first step, we discussed what AI can do today in many different areas, including computer vision, natural language processing, speech recognition and other fields. We narrowed down to a small project in computer vision, identifying objects in images near real time, that can be plugged into the current software.
Project Evaluation Criteria
The project is related to the core offering, and solves a real problem, even that it does not provide a huge financial impact right away.
The project is feasible, meaning that it can be implemented end to end, from picking up the image via the camera, to output categories of the object feeding back into the software. It is also feasible as we know it will achieve a good accuracy, although optimize accuracy is not the priority at this stage.
The project is quick to start and can be completed within a short time frame. It mainly does not depend on large input from multiple departments from the Client.
AI Algorithm Development
Gravitate AI developed the initial AI algorithms using TensorFlow and transfer learning. We "wired" all the steps together, utilizing a combination of the open source TensorFlow Object Detection API and proprietary code. Different model versions were tested to compare pre-trained models hyper parameters. Algorithm results were compared from both computing efficiency and accuracy perspective. The final algorithms were recommended to the Client.
The "starter" source code were provided to the client. The algorithm was implemented by the Client's software team via docker using TensorFlow serving. The restful API takes real time encoded images, and outputs JSON files with object coordinates, object categories and their associated probabilities. The output files then fed into the overall software solution to build more comprehensive features delivering value to the end users.
Training was provided to both of the business team and technical teams. For the business team, we focused on industry practices, and how AI can provide value to the specific business. The internal technical team learned about object detection literature, deep learning basics, "starter" code walk through as well as project presentations. Gravitate AI continue to support and train the teams in AI and data science subjects. Click here for a Gravitate AI training deck.