Cerebry - AI driven math teaching assistant

Cerebry is a virtual math teaching assistant that works with existing math curricula and accelerates student mastery of concepts.

Practice is key to understanding concept-based subjects, such as maths. Textbooks and free online videos can help with the initial understanding, but they do not reinforce concepts or create lasting proficiency for exams. While some students have access to private tuition to help, the vast majority do not.

Cerebry solves this problem with a web-based, AI driven, virtual tutor that:

1. generates questions at individualized difficulty for each student, for levels ranging from primary school to tertiary education 

2. provides interventions like hints, step-by-step solutions and breakdown of complex problems - just like a “human” tutor would do

3. provides unlimited practice with new original questions being AI generated until the student achieves mastery.

Born out of research at the National University of Singapore, Cerebry’s AI engine can encapsulate the rules of a well-defined domain as axioms, much like a human teacher. This allows the engine to generate questions on-the-fly, solve them and diagnose a student’s precise misconceptions.

Students attempt questions similar to how they would with a textbook (a basic smartphone with low bandwidth connectivity suffices). They select an exercise and start answering questions. Problems automatically adjust to the student’s proficiency. Teachers meanwhile can see exactly where students are struggling and where they can intervene at a learning objective level to accelerate the process of attaining mastery.

Many private schools in the Philippines (following the local DepEd curriculum) are actively using Cerebry, already in their second year. Some have stopped ordering textbooks and rely solely on Cerebry instead. In studies we have demonstrated that the use of Cerebry improves learning outcomes significantly. We know that by deploying our solution more widely, namely across public school systems, currently underserved students would benefit even more, thus contributing to closing the gap between private and public education.

Numeracy levels across most of the Octava Social Innovation Challenge’s target countries, with the exception of Vietnam, are lacking significantly as evidenced by the low reported PISA scores, culminating in the finding that less than one-third of 15-year old students in the Philippines achieved a minimum proficiency level in maths.

There are many reasons for why this is so. Insufficient teacher education, large student-teacher ratios, high numbers of children out of school; the list goes on. But the core problem is that in all these countries, government funding towards education is inadequate. To put this into context, Singapore’s Ministry of Education’s (MOE) 2020 budget was 9.3 billion USD for ~500,000 students. On the other hand, the Philippines' Department of Education (DepEd) had a slightly higher budget at 9.8 billion USD to serve 26.2 million students!

While Cerebry cannot solve the funding issues, we can help the existing funds and teacher resources to go further by complementing and multiplying (and sometimes substituting) math educators’ contributions, through our adaptive, personalized teaching assistant. By equipping public school systems nation-wide with our solution, we can create equitable access at scale.

While students and teachers from all backgrounds benefit from Cerebry, low resourced students and under-resourced teachers stand to benefit the most. 

Low resourced students will be able to compete with high resourced students, opening doors that would not be open to them, without our support. Dedicated students will have the greatest marginal benefit and that could mean life changing opportunities, such as admission to universities, which otherwise would not have been attainable. That said, even less dedicated students will build foundational math skills, which will avoid them to be left behind, and provide them with the necessary tools to function in today’s world. 

Underresouced teachers will have insights they could not have imagined before. They will have the ability to see exactly which concepts have been mastered by their students and which concepts need reinforced. In the past, teachers may have been able to spot glaring weak areas, but nowhere close to the granular feedback the Cerebry provides them.  This insight will allow the teachers to utilize their time more efficiently and ultimately impact PISA scores.

Poor performance in math is almost universal across the Octava Social Innovation Challenge’s target countries (with the notable exception of Vietnam), while our solution has been demonstrably able to improve our learners’ level of understanding in maths.

Equipping public and low-cost private schools with our AI driven maths teaching assistant thus directly aligns with Octava’s Challenge, as this would bring an affordable, as well as accessible, EdTech solution. A solution that directly aligns with each country’s defined curriculum outcomes to a breadth of K-12 learners, currently studying in frequently underserved learning environments.

Our low marginal cost, once a curriculum is represented in the system, enables us to provide our service at very affordable rates (possibly for free upon engaging potential partners). Moreover, thanks to our web-based delivery, students and teachers only need access to a basic smartphone, along with low bandwidth internet access, to use our solution.

Shubham Goyal, CEO and co-founder of Cerebry

Cerebry does not rely on question banks and decision trees. Instead, questions are generated with artificial intelligence, via deductive reasoning from a knowledge base represented in first-order logic concepts. Generated questions that satisfy criteria such as concept fit, curriculum fit, and required level of difficulty are then processed by a natural language generator resulting in the actual question (and the break-down solutions) presented to learners. This novel approach is the result of the founders' PhD thesis at the National University of Singapore (NUS). 

If a student gives a wrong answer, Cerebry not only solves the problem step-by-step for the student, but generates a new similar problem on-the-fly. If still struggling, the engine can break down the question into its component concepts or give hints. Once the engine diagnoses exactly where the student is struggling, it starts to remedy that issue by giving the student tutorials and generating practice questions for that skill or bringing the difficulty level down. This level of personalisation is simply not offered by any other adaptive product which explains the low level of penetration of adaptive learning.

In summary, the benefits over traditional approaches are three-fold:

1. First, as questions get actually generated on the fly, virtually no two questions are the same, they get generated covering the right set of concepts at the right level of difficulty in relation to the desired learning outcome and the learner’s current proficiency. Plagiarism is a non-issue.

1. Second, ability to process learners’ answers enable the engine to autonomously detect mis-conceptions and in truly adaptive fashion, hone in on those mis-conceptions via break-down solutions, break-down questions. The engine also “remembers” previously encountered mis-conceptions and repeatedly validates those concepts’ understanding with learners.

Third, this approach scales significantly better than traditional template-based question and solution generation, as only generation parameters, course structure, and language need to be configured to adjust to a new or revised (national) curriculum.

Cerebry’s current maths teaching assistant was launched in 2020 for mathematics. We have already been able to cover several curriculums - Philippines DepEd, Cambridge IGCSE, O Levels, A Levels, CBSE, IB etc. at minimal cost.

In the Philippines, Cerebry is being used by private schools to improve student outcomes. Teachers began with using it as an additional tool, but now it has completely taken over homework and even assessments. We have conducted surveys with students in several of these schools, >90% of students report feeling more confident and even report that Cerebry helped clarify misconceptions they did not know they have.

Singapore University of Technology and Design (SUTD) published a peer-reviewed paper measuring the impact of Cerebry and Vedantu, an online tuition platform, conducted a month-long trial of Cerebry and found that students using Cerebry had a 37% improvement in their post-test scores compared to the pre-test.

The core of Cerebry is powered by two enabling technologies:

(1) Autonomous Question Generation

Cerebry does not use a question bank. It represents the domain knowledge of any well-defined domain as first order logic axioms. The addition of domain knowledge closely mimics a teacher teaching new topics to a student. The engine uses deductive reasoning and constraint handling rules. The codebase is written in Prolog. The adaptive layer would pass the following parameters to this layer to generate a new question:

• Concepts to be included in the new question

• Skills to be included in the new question

• Overall difficulty level and difficulty values wrt each concept and skill

• Student’s competency matrix

• Curriculum and grade

This layer not just generates the question, but also the step-by-step solution, conceptual breakdown, hints and other interventions which the adaptive layer might need to provide for the student. Apart from this, it also provides metadata to the adaptive layer to aid in its decision making process.

More details on domain knowledge representation and a simple algorithm to generate questions can be found here: (https://www.researchgate.net/publication/300790066_A_Framework_for_Automated_Generation_of_Questions_Based_on_First-Order_Logic).

(2) Adaptive Layer

This layer decides what questions or interventions the student gets presented based on its assessment of the student’s proficiency. It invokes the question generation layer when needed.

Recognising the difficulty of items is essential for a completely autonomous system. Perceived difficulty based on student responses is one of the factors, and not the one with most weightage. Other important factors and how they are used internally are described here: https://ieeexplore.ieee.org/document/7814689.

There a few factors that feed into its decision making:

• Student’s last response (item response theory)

• Pedagogy chosen by the school

• Target proficiency

• Topic and subtopic

• Student profile (motivation, inferred learning style, etc.)

At Cerebry, our overall mission, to “Democratize Education”, is being accomplished with our internally developed AI engine and our virtual tutor. Our Theory of Change is largely consistent with the five solution areas of the Octava Social Innovation Challenge.  We have incorporated the Octava solution areas into our Theory of Change.

In line with our Theory of Change and based on our already proven efficacy of the solution, the input metrics towards achieving our desired impact are mostly based on access, use, and engagement:

1. Active curricula coverage in maths, for the five Octava Challenge countries, currently stands at 8% of the student population. The barrier to progress is mainly language translation and we are confident that we can grow our curricula coverage.

2. Active student user base (measured in daily, weekly, and monthly active users) - in absolute terms as well as in relative terms to the student population. For reference, we currently stand at ~10k monthly and ~6k weekly active users, in Southeast Asia.

3. User engagement (measured by session duration and number of questions attempted per session).

4. Academic impact - we plan on collaborating with schools and governments to further solidify the causality between practice on Cerebry and better performance (exam results) in maths.

Secondary progress metrics are more qualitative in nature and aim at increasing awareness and accessibility as a precursor to usage:

• Number of scientific studies proving our approach

• Number of relevant endorsements from education bodies

• Number of distribution and/or financing partners

The most objectively measurable, high level impact goals of Cerebry, relate to positive changes in PISA results, which will have transformational impacts on the lives of students in the five focus countries of the Octava Social Innovation Challenge. As we are confident in the efficacy of Cerebry (discussed in the prior questions), achieving our high level impact goals is a matter of rapid and widespread adoption by schools, and usage by students, of our platform. 

In one year from now, we expect to have made significant inroads with the governments and schools in the Philippines and one other country (e.g., Indonesia). In three years from now, we expect to be fully accredited by the government in the Philippines, Indonesia, Thailand, and Malaysia. And in five years from now, we aspire to achieve more than 50% penetration of the student population, in all public schools, in the five focus countries of the Octava Social Innovation Challenge, either by direct operations or by having integrated with EdTech partners in those geographies.

To date, technology innovation has had the impact of widening the gap of education outcomes, between low resource students and high resourced students. That said, a platform like Cerebry, which is highly scalable, at a low cost per student, will rapidly impact lower resourced students and is poised to reduce that gap.

Access to and coordination with government and non-government institutions supporting the public school system(s) in our target countries, as well as the ecosystem such as official textbook providers, publishers, etc.

The key barrier in disseminating our solution into public education systems in the Octava Social Innovation Challenge’s target markets lies in the actual go-to market. Specifically, in generating interest and sufficient commitment from the public (or low-cost private) school systems, or their backers, to enable the development / configuration of our solution to fit the local syllabus and language.

In absence of such up-front commitment, a smaller hurdle lies in the cash outlay required for adapting our solution to the national curricula (including the effort for translation, quality assurance, etc.).

As a “school by school” approach doesn’t scale well, our plan sees us collaborating directly with national school boards and public education service providers. We would integrate, both technically and in outreach, with their existing platform. The MoEC platform in Indonesia, as well as DELIMa in Malaysia, are such examples.

Where such platforms don’t exist or are not in widespread use, for instance in the Philippines, we hope to be able to work with nation-wide publishers, serving government schools.  Also, we plan to address public school teacher networks directly, for instance by working with Philippine Normal University, which organizes public school teacher education for the whole of the Philippines. Networks of private low-cost schools (e.g., APEC schools in the Philippines) provide an additional avenue.

Cerebry started as a PhD project by Dr. Rahul Singhal (https://dblp.org/pid/42/7798.html) at the National University of Singapore (NUS) in 2011. He set out to create a virtual teaching assistant that can assist millions of students in developing countries who cannot afford expensive private tuition. Freely available content was only a part of the solution, but not the complete answer since practice and tutoring are essential to crack standardised tests. Rahul believed that human-centered solutions would always be expensive and never available to students who are outside metros, therefore an AI-based inexpensive personal tutor was the only solution.

Shubham met Rahul during the course of his PhD and they built the framework together during the latter part of Rahul’s PhD. Both of them came from humble backgrounds and had to move to metro cities for the purpose of preparing for engineering entrance exams. This problem affects millions of students, and most cannot afford to change where they live. They formed Cerebry in 2016.

Cerebry raised a small round of funding and began work on teaching the engine the domain of mathematics. Cerebry was awarded an author’s contract for multiple workbooks beating human authors and that validated the vision proving a completely autonomous system was not just possible, but comparable to humans.

Cerebry raised two rounds of funding in 2018 and 2019 for further R&D. In 2020, Cerebry deployed the first product in schools in two countries - Singapore and Philippines. Adoption immediately kicked off and Cerebry became the platform of choice for the top private schools in the Philippines. Universities in Singapore started relying on Cerebry for all their practice and assessment needs.

Cerebry raised another round of funding in 2021. Cerebry now serves over 15,000 students in 4 countries.

Currently we have 126 employees and advisors, up from 75, one year ago. Our team members are located in India, Singapore, Philippines, the United States and Australia.

The team at Cerebry is a mix of AI researchers, software engineers, curriculum experts and teachers. Cerebry’s current headcount stands at 126 with a particular emphasis on technology (59) and learning outcomes (55).

The management team includes

1. Shubham Goyal, co-founder and CEO (https://www.linkedin.com/in/shubham94/). Shubham has led technology teams at other start-ups like Holmusk and Dexecure and is an engineer himself.

2. Dr. Rahul Singhal, co-founder and CTO (https://www.linkedin.com/in/dr-rahul-singhal-1134674/) - Rahul leads the AI team and was the creator of this framework.

3. Rohit Singhal, co-founder and Director of Engineering (https://www.linkedin.com/in/rohit-singhal-0453a4a/) - Rohit was the CTO for large brick-and-mortar tuition centers in India and led the transition of adda24x7 from completely offline to a mostly online (>70% of revenues) model

4. Christian Nill, CCO (https://www.linkedin.com/in/christiannill/) - Chris started his career for Boston Consulting Group (BCG) and then joined Alibaba. He has led product, sales and marketing teams in his 15 year long career.

5. Vei Li, Head of Content - Vei Li has more than 25 years of experience as teacher, head of department and curriculum planning at the Ministry of Education (MOE), Singapore. She has taught multiple curriculums in Singapore and Australia. She is a widely published author of books on mathematical problem solving.

With such a diverse and experienced team, Cerebry is well positioned to bring this game changing innovation to the five target countries.

The founders, management team, and angel investors of Cerebry are driven by a mission to improve maths education for all members of society. At the same time however, in order to sustain the considerable operations that we have set up, we needed to focus on the commercial success of our young company. This means that we currently market our solution to the more affluent segments of society – in line with the findings of the Octava white paper, with regards to the existing focus of EdTech companies in Southeast Asia.

Yet we aspire to be used in public schools all across Southeast Asia (and beyond) as we believe Cerebry provides a unique opportunity to improve the quality of maths education, at a low cost.

Octava is uniquely positioned to help us democratize what we have built, by way of establishing Cerebry as a tool used in public education.

We will use the grant money received through this competition to extend our coverage of national math curricula in the region, starting for instance with the Indonesian curriculum in Bahasa Indonesia, according to the syllabus set by the Indonesian Institution of National Education Standards.

We will work with the Octava foundation and its partners to establish partnerships with government and non-government actors that help us ensure sufficient scale in our deployments to schools in order to safely recuperate initial investments, while keeping the cost to schools as low as possible.

In particular, we can collaborate with other local/regional EdTech players, such as Ruanguru, Zenius, or FrogPlay. By acting as “back-end service” for those other EdTech companies, we could greatly increase the reach of our adaptive practice.

First, we would like to engage in defining country-specific business models, to specifically service the public education sector, in line with the Octava challenge. We’re very open to refining our business model(s) to better cater to the “bottom of the pyramid” user base.

Secondly, as we extend curriculum coverage to national curricula in local languages, tapping into Octava’s network can help us in recruiting for positions in curriculum and content development - extending our existing AI engineering team.

Third, and likely most important, support in building out a network of distribution and service delivery partners across government and non-government players, in the target markets, can help us achieve scale targets.