We develop education-specific LLMs with enhanced reasoning capabilities: Vineet Govil, CTPO, PhysicsWallah

In a recent interview with Express Computer, Vineet Govil, CTPO, Physics Wallah (PW), shares insights on the rapidly evolving edtech industry in India. He discusses the transformative impact of AI and ML on personalised education, highlighting PW’s initiatives like the AI Guru and Smart Doubt Engine. Govil emphasises the importance of reaching students in Tier 3 and Tier 4 cities and speaks about the potential of AI in revolutionising education, from automated grading to personalised career guidance. He also addresses challenges such as AI’s reasoning capabilities and the need for emotional well-being support for students.

What is your perspective on the current scenario of the edtech industry in India?

In my opinion, the edtech industry in India has evolved remarkably. Although technology was used in education before, its popularity surged during the COVID pandemic, which acted as a catalyst. With schools closed, online classes became essential for students to continue their education. Since then, the industry has progressed significantly.

Initially, online classes during the pandemic involved basic setups. However, post-pandemic, advancements in AI and ML have brought substantial customisation and personalisation to education. These technologies have enabled more effective learning experiences for students.

There are three key stakeholders in this ecosystem: students, teachers, and parents. While students are the primary focus and teachers are crucial, parents also play an essential role.

Students in grades 5 to 9 spend significant time at home, and while parents may not directly teach academic subjects, they contribute to core values important for a child’s development.

At PW, we involve parents by ensuring they understand their children’s activities and support their learning journey.

What benefits do you see in incorporating GenAI into this model for both you and the students?

GenAI helps create content more quickly and effectively. At PW, we integrate it responsibly, ensuring that our technology serves our society and students well.

Our AI suite is available in PW app on both mobile and web platforms. Earlier this year, we launched Alakh AI, which includes several products like AI Guru. AI Guru is a 24/7 tutor available for any academic questions, accessible via text, images, or voice. It uses our rich content library, providing answers with detailed text explanations and relevant videos from our repository, not generic online sources.

For example, if a student asks about Newton’s first law of motion, AI Guru directs them to the exact moment in our video library that covers the topic in depth. AI Guru also addresses non-academic questions. There are several motivational videos available on our app to help students who struggle with focus or motivation, which we provide to support their needs.

Is your AI model based on an open-source framework or a custom-built solution?

We are using the OpenAI’s GPT4o & GPT4omini models. However, as we know the LLMs can hallucinate, we’ve ensured the answers are grounded to the student’s context. Technically, we’ve built an agentic RAG (Retrieval-Augmented Generation) that detects students’ intent and utilises the proprietary content we indexed into our vector database. Thus, this enhances students’ confidence as the information comes from the trustable resources.

Do you have insights into whether your student demographics are predominantly from metro, tier-1, or tier-2 cities?

Our vision is to provide quality education at a very affordable price to students who cannot afford expensive studies. We genuinely want to ensure that we reach out to Tier 3 and Tier 4 cities, along with, of course, the metros. The students in these areas are often deprived of the support they need. They can’t afford expensive education, and they don’t have many alternatives.

Currently, we’re present in almost 98% of the pin codes in India. However, if I had to break it down demographically, Uttar Pradesh, Bihar, and Maharashtra. These states are leading in terms of volume of traffic. We’re also seeing strong growth in the southern states, particularly in Karnataka and Kerala. In Kerala, for instance, we made an acquisition through Xylem, which is now part of our portfolio. We’ve completely revamped that platform, which has also contributed to the increased traction there.

Are you focusing on upskilling students to be job-ready and industry-ready with skills in GenAI?

Yes, we do have a separate vertical dedicated to upskilling. We offer professional courses to students, where we ensure they learn about GenAI. Additionally, even in our foundation and pre-foundation classes, which cater to students from 5th to 8th grade, GenAI is included as one of the subjects.

We teach them how to use it effectively. The beauty of our approach is that we cater to a wide range of students, starting from as young as 2nd or 3rd grade, up to 12th grade, and even professionals.

So, do you think there still needs to be a change in mentality among consumers?

There’s a belief among some parents that online education might not be as effective as traditional methods. However, others believe that online education can be just as effective. To address this, we offer both online and offline options. Our online approach includes daily practice assignments and live classes tailored to the student’s stream and grade. These live classes are also available as recordings for offline access.

We previously discussed the AI Guru, but another key tool we have is the Smart Doubt Engine. This tool addresses concerns that online education might not handle student questions promptly. The Smart Doubt Engine allows students to ask questions during live classes, with immediate responses provided by the engine for most of the questions.

Before launching this technology, we conducted extensive testing with our team of teachers to ensure the accuracy of the machine-generated responses. Only after validating the responses did, we deploy the Smart Doubt Engine.

What are the current challenges that AI, particularly GenAI, faces in improving reasoning capabilities, and how are you addressing these challenges?

Yes, there are many challenges, but if I have to mention a few, one significant issue is that AI, particularly GenAI, still struggles with reasoning problems. Take any state-of-the-art Large Language Model (LLM). We are also collaborating with some big names that I can’t disclose right now. We work with these companies to define how to modify or create education-specific LLMs with better reasoning capabilities so that the accuracy of our responses can reach the 90% range, which is very difficult to achieve today.

Additionally, we’re tackling another interesting problem, we have AI suite that includes multiple products. I’ve talked about AI Guru and Smart Doubt Engine, but there’s one more product that’s very close to my heart: AI Grader.

This tool has been launched for the UPSC category. It analyses handwritten subjective answers and provides feedback to students on what they could have written better, where they went wrong, and how they should have structured their answers. I believe this has enormous potential for the future. Imagine if AI could correct exam papers—be it for CBSE, ICSE, or any other board—this would revolutionise the education landscape, not just in India but worldwide.

Do you have human oversight to ensure the accuracy of these tools?

Yes. As I mentioned with the Smart Doubt Engine, before we launch, we ensure that each answer we provide is vetted by a human subject matter expert (SME). The same rigorous testing and accuracy checks were done before launching the UPSC category feature. We made sure that what we offer is reliable and correct. However, it’s important to note that, as with any machine, there can be errors.

Have you experienced any specific hallucinations with these systems?

Yes, we have, especially during beta testing. We’re very open about this with our student communities enrolled in the beta program. We inform them that there might be instances where the responses aren’t correct. There were indeed situations where we didn’t provide the right answers. We took feedback, corrected our systems, and incorporated those improvements.

Any final thoughts about the development of the tech industry and PW?

In my opinion, the edtech evolution we’ve seen in the last two to three years is just the tip of the iceberg. There is immense potential in the future. For instance, entire exams and notebooks could be corrected by machines. Personalisation could reach a level where, instead of a physical teacher in schools, students could have their favourite teacher—let’s say, for example, Alakh sir—appear through virtual reality, teaching them as if they were standing right in front of them.

Additionally, GenAI is progressing rapidly, with improvements in accuracy and personalisation. Customisation could eventually become so advanced that, for instance, a student preparing for the JEE exam could be advised by AI that they might be better suited for commerce, NEET, or another field. The AI could guide them in their career choices and where to focus their efforts.

I’m very optimistic and excited about the future of edtech. As I mentioned, this is just the beginning; we have a long way to go. Another important issue, which we didn’t touch on earlier, is the emotional well-being of students. Sometimes students feel low or break down if they can’t perform well, often due to parental pressure to achieve certain marks or clear specific exams. Some students can cope with this, but others struggle. There needs to be technology available to help students navigate these challenges and come out stronger.