Conversation with Fuse Classroom’s AI PhD

AI has had a tremendous impact on the teaching and learning process. From the automation of tedious tasks to personalized learning, AI-powered platforms such as Fuse Classroom are revolutionizing education.  

To learn more about how AI engineers integrate AI into Fuse Classroom, we spoke to Fuse Classroom’s Manager of AI development, Mr. Rakesh Katuwal. We discussed what it’s like to build AI models and integrate them into LMS platforms, and what daily challenges he and his team face while testing these models. Here’s the full conversation.

Tell us a bit about your journey of becoming a Manager of AI development

After finishing my PhD in Machine Learning, I started working as an AI researcher at Nanyang Technological University, Singapore. At that time, I was more into AI academia. I came to know about Fusemachines and how the company is at the forefront of democratizing AI through education and solutions. After doing research about the company and connecting with the company’s Founder and CEO Mr. Sameer Maskey, I was inspired by the company’s vision and mission to make AI accessible to underserved regions. So in March 2020, I came back to Nepal and joined Fusemachines as a Senior ML Engineer. During my first few months, Sameer guided me a lot. He always believed I was capable of becoming an industry expert. Honestly, I didn’t think I would be here if it was not for Sameer’s beliefs and guidance. Eventually, I undertook more and more projects and began managing the whole team. Now I’m leading the Fuse Classroom AI features development and integration process.

What’s a typical day for an AI Developer?

My day starts with team meetings and stand-ups. We have multiple squads involved in different AI model development, testing, and deployment processes. As the Manager of AI development, I make sure to touch base with each squad and get updates related to any progress we’ve made on the platform. Most of my days range from discussing problems related to issues in teaching and learning that AI can solve, ideating new features in which AI can be leveraged, and collaborating with multiple teams — particularly Product, Data, ML, QA, and Backend — to create and build on AI strategies. I try to plan out my work schedule before I start and it has helped me stay focused and organized.  

Can you walk us through the process of designing an AI system?

There are fundamental steps to keep in mind while designing an AI model. The first is Problem Identification where we try to answer two important questions: (1) What are we trying to solve? (2) What is the desired outcome? For this, we coordinate with the Product and Customer Success team to identify problems that need solving. After defining the problem, the next step is Data Collection. We need to gather all the relevant data and label them in order to train the AI system we want to build. This is one of the most important steps, as without relevant data even the most ingenious algorithm will be useless. 

Following this, we choose the most appropriate algorithm, train and deploy it. The final step is monitoring and evaluating the system based on how well it’s performing and the impact it’s creating for users.

What day-to-day challenges do you face while testing AI models in Fuse Classroom?

One frequent challenge we face while deploying AI models is how the model predicts the output. Lack of relevant or “good” data can create AI systems that are biased. An AI bias is when an algorithm produces prejudiced or biased results due to inaccurate assumptions in the machine learning process. For instance, Google’s first generation of visual AI identified images of people of African descent as gorillas. We obviously want to avoid all bias. The next challenge is data scarcity. Because platforms such as Fuse Classroom require students’ and teachers’ test data, their behaviors, courses and curriculum, it’s a big struggle for us to build an AI model without relevant data to solve a particular problem catering to a particular user. That’s why we are still gathering as much data as possible and training them. 

And the final challenge is end product validation. We need to validate and verify whether the AI feature we created is solving the problem in the right way. A constant evaluation and reiteration of every step mentioned above is required when it comes to validating the end product and providing the best user experience to our stakeholders.

Despite the constraints, we have managed to build advanced and useful AI features such as an AI proctoring tool, AI recommendation system, student status engine, and many more in Fuse Classroom and the feedback has been positive.

What are the best goals to keep in mind when creating an AI strategy? 

Building an AI model is usually problem-centric, but in my experience, a successful AI strategy should be user-centric. It should create value for end-users. When it comes to solving a real-world education problem, the first goal is thorough research of what’s happening in the education industry. The next goal is to select the most pressing problems that need immediate action. Following this, another goal is having a data strategy. We should collect as much good data as possible to train the model accurately. And finally, we need to define the vision of the AI features. We need to keep in mind internal glitches such as lack of structured data, proper product validation and other challenges around bringing our AI projects to life before creating a successful strategy. Based on all these insights, the product team and AI team brainstorm to create an AI product roadmap. 

A successful AI strategy should be user centric. When it comes to solving a real world problem related to education, we need to research the current scenario, empathize the problems that really needs solving and finally define the vision of the AI feature.” – Rakesh Katuwal

Your team is working on many AI tools for Fuse Classroom. Among those, for which tool has your team had to run the most tests?

Different features in Fuse Classroom have been integrated with different AI tools, so every feature has been tested and changed multiple times to get the results we want. The Student Status engine in Fuse Classroom, which uses a simple ML model to predict the overall performance of every student, has been tested multiple times. The engine works by alerting teachers with color-based indicators when a student is either doing well (green), average (yellow) or poorly (red). The unexpected challenge we faced with this engine was validation and trust. Most instructors initially disagreed on criteria that defined which student should fall under which category and we had to work to reiterate and find common ground where the teachers were satisfied with the insights they got from the platform.

What’s the role of AI in the education industry and how is Fuse Classroom affecting change?

AI is disrupting every sector including education. AI-enabled LMS are booming at times of crisis like this. Fuse Classroom is a unique platform that leverages artificial intelligence to automate tedious tasks, save time and costs and help schools improve the learning experience. The platform can be customized as per the school’s needs and consolidates features such as LMS, SIS, live class, admission, graduation, online exams, and school community together in one platform.

Fuse Classroom is set to reimagine the classroom by assessing student progress, recommending study material, and including a chatbot that assists users with the platform making learning easy, independent, interactive and enjoyable.

What’s a piece of advice you would give to other AI developers?

One of the most common problems with AI developers that I’ve seen and experienced myself is starting out with a complex algorithm and programming technique. There are so many AI tools, libraries, and so on to build prototypes and engineers have high hopes while building AI products thinking it will solve a big problem. But in real life, it’s not as easy and you have to understand the basic concepts first. You need to have a good enough data strategy and focus more on validation and user experience rather than calculations and numbers.