Kasuku Stories

There are more than 250 million school aged children in the world who don’t know how to read or write, many of whom don’t have access to schools or teachers at all. UNESCO estimates that addressing this shortfall would require an additional 69 million teachers by 2050.

Through a combination of story-based learning, social software, and machine learning techniques we call “digital personalities”, we build technology that enables children to teach themselves basic literacy and numeracy skills in a way that can be applied across a wide range of languages and cultures. 

We hope that by empowering children and families to teach themselves basic literacy outside a traditional school context, we can help close this instructional gap and help ensure that the next generations of children have the opportunity to be literate.

UNESCO estimates there are over 250 million school-aged children in the world who do not know how to read, write, or perform basic counting. Many of these children do not have access to schools or teachers at all, and UNESCO estimates it would require an additional 69 million teachers to address this shortfall by 2050. Even with an unprecedented investment in early literacy instruction, we would still be left with entire generations of children without the opportunity to become literate.

This is not merely a problem in the developing world. The US 2017 NAEP Reading Report Card identified 32% of grade 4 children at “below basic” proficiency in 2017, a number that has not changed significantly for 10 years. Furthermore, data from the Annie E. Casie Foundation suggests that early intervention is essential, as children who are not reading proficiently by the end of third grade do not graduate from high school on time at rate four times greater than that for proficient readers.

Kasuku Stories is focused on addressing the early literacy gap, specifically, by helping preschool-aged children achieve basic literacy skills outside the traditional classroom.

Kasuku Stories is  aimed at children aged 3-6 (preschool-aged children) who would benefit from basic literacy and numeracy instruction delivered outside the context of a traditional classroom. This may include disadvantaged populations, low-income learners, indigenous populations,  children in developing nations, immigrant populations, ESL learners, and home schooled children.

Kasuku Stories was initially developed as an entry into the $15M Global Learning XPRIZE (GLEXP) challenge. The original version, an Android version developed in English and Swahili, was first tested with a small group of children aged 4-9 at the the Mercy and Caring children’s home in Kitale, Kenya, and later, with school-aged children in Tanzania. The original English version was field tested in Ottawa, Canada in after-school Kindergarten programs (3-5 year olds). In 2017, Kasuku Stories was named one of 11 semi-finalist for the GLEXP.

Since then, we have been adapting Kasuku Stories for distribution through the Apple and Android app stores, and beta testing across English speaking countries in North America, Europe and Australia to better understand the needs of preschool-aged learners in the various demographics described above.  

There are three key innovations that set Kasuku Stories apart from other technology-based literacy solutions: story-based learning, social software, and digital personalities.

In order to accommodate a wide variety of learners, Kasuku Stories is designed to be adaptable across multiple languages and cultural contexts. What’s more, it’s designed to be used in environments outside the traditional school or classroom. How do you encourage a child to want to learn if there’s no one to tell them they have to do it?

“Story-based learning” is our approach to  motivating, scaffolding, localizing, and scaling our literacy curriculum. In story-based learning, all of the learning activities automatically adapt to whatever stories the child is reading. This allows us to adapt the curriculum to a new language or culture simply by changing out the set of stories that are available.

The second innovation is “social software”: software that encourages children to work together on the device. Often, when technology is introduced into an educational environment, it is with the goal of one device per child. We believe this to be a missed opportunity. By encouraging multi-child use of a single tablet, we can foster a much richer learning environment, and make technology-based learning more accessible to learners. By enabling children to work together, encourage one another, even teach each other, we expand learning opportunities beyond the confines of the virtual world.

Wherever possible, our software and learning activities are designed so that many children can learn together simultaneously. Though this introduces some interesting design challenges, the resulting group learning environment is a wonderful social reinforcement to the learning. By enabling children to work together with their friends, siblings and parents, we make learning a fun, social activity. This increases the likelihood that the learning activities will become part of their daily lives.

Finally, we incorporate machine learning to help make pedagogical decisions. We call our  blend of technologies “digital personalities.” These personalities serve three main purposes. First, they maintain engagement by suggesting content that the child will most likely enjoy (modelling "preferences"). Second, they ensure that literacy curriculum is covered in a pedagogically sensible way (implementing “curriculum”). Finally, each tablet maintains its own latent set of preferences and behaviors (exhibiting “personality”). In an environment where multiple tablets are available, this leads to self-selection among the children, with children choosing the tablet which bests suits their style of learning.

One of the innovations of our “Story-based learning” approach is that our curriculum will automatically adapt to the set of stories loaded onto the device. This makes it easy to localize the software to a new language or culture, simply by exchanging the story set. This greatly improves our ability to scale and adapt to new learning environments.

Our "Digital Personalities" employ AI techniques drawn from the field of  Intelligent Tutoring Systems (ITSs) to learn child preferences, and maintain progress through a curriculum. Furthermore, when multi-device deployments are used, devices can exhibit distinct “personalities,” permitting children to self-select the personality (device) that best suits their level and learning style.

Social Learning: Our user experience (UX) is designed to encourage multiple children using the device at once. This permits learners to become teachers as they grasp new concepts.

There are two novel technology components that underlie our solution: AI, and UX:

At the core of our software (the “Digital Personality” component), we make use of several AI technologies:

• A general (Bayesian) decision engine, for learning user preferences and exhibiting device personality, and

• A pedagogical engine, to make decisions to balance user engagement, and progress through a curriculum.

To train these engines, we make use of a multi-agent simulation. A User Agent is trained on usage data and makes it possible to simulate vast numbers of interactions with the software. Preference and Pedagogical Agents are then trained using a variety of machine learning algorithms (e.g. Bayesian Knowledge Tracing, Reinforcement Learning) until the desired characteristics are learned.

In addition to machine learning, we also focus on User Experience (UX) innovations. In particular, we are developing UX innovations aimed at encouraging multiple children to use a device at once. This can be a surprisingly hard problem in many application domains. Even a basic drawing app can be difficult to scale to multiple users simultaneously (e.g. how can one user choose something like line color without affecting the line color of another user?)

There are 250 million school aged children who do not know how to read or write. If we are going to build literacy in these children, we need to begin by identifying where the current systems are failing them.  Many of these children are not in school at all. The rest are somehow under-served by the current educational framework.

Our solution is designed to supplement teacher-delivered education in contexts where it isn’t available, or isn’t sufficient. We know this is possible. We began as an entry in the Global Learning XPRIZE grand challenge competition. Every one of the field-tested finalists in the GLEXP was able to demonstrate literacy gains in a target population.

We aim to improve on existing solutions by targeting three key areas: engagement, relevance, and personalization. Story-based learning achieves several of these ends by engaging the children with relevant, culturally appropriate content. Social software encourages learning in a community, which further enhances engagement and enables learners to teach each other. Digital personalities help personalize the experience, balancing engagement with progress against a curriculum.

We have used opportunities like XPRIZE competitions to help vet, test, and validate our approach. We were a semifinalist in the Global Learning XPRIZE, and one of 30 solutions to proceed to round 3 of the the (still-active) IBM Watson AI XPRIZE. Now we are bringing out solution to the mass market, and MIT Solve can help us achieve the scale and reach we need to tackle the global literacy gap.

We currently serve 100 users (our initial beta test population).  We launched in the iOS app store in April, and are actively growing our user base. We intend to serve 1,000 by the end of 1 year. We are aiming for 10x growth per year, and thus we are aiming to reach 10 million children by the end of 5 years.

Validate our business model. In April, we launched our first product in the iOS app store as a subscription product aimed at the North American market. We need to validate this part of our business model, at least for the developed world.

Refine our AI models. We are constantly looking for ways to improve our learning outcomes. By the time the AI XPRIZE wraps  up (2020), we aim to have validated our high-level machine learning  concepts that support our pedagogical engine. The various subfields of AI are developing rapidly, however, and there are many promising areas of deep learning we wish to explore.

Add Global Development Deployments. We aim to seek out partners for potential white label deployments in the developing world. (We do not have the expertise or resources to design and implement ethical and effective interventions in the developing world on our own at this point in time).

Develop our content. To this point, we have built a framework that is largely independent of the content we put into it.  The vast majority of our content has come from open content providers such as the African Storybook, Book Dash, and StoryWeaver projects.  Going forward we need to develop a more robust content development (or content partnership) strategy.

Add additional languages, multilingual support. Our software was initially developed with English and Swahili versions. Next we wish to add French and Spanish, as well as mixed-language (bilingual) support. This will greatly increase our solutions usefulness among ESL learners.

Market Challenges: Our number one challenge is whether we can find a sustainable and profitable business model.  The preschool educational technology market is crowded, and while we feel we have significant technical advantages, differentiation in this space may be difficult. We need to find ways to reach our target customer, and that customer must be willing and able to pay for our solution.

Cultural Expertise: We do not have the expertise or resources to design and implement ethical and effective interventions in the developing world on our own.

AI Technologies: the pace of change in the AI space can only be described as frenetic. We likely do not have the resources to keep up on the technological advances, nor fund large-scale machine learning experiments that will be necessary to stay on the leading edge.

Cultural Expertise: We originally participated in the Global Learning XPRIZE as a means of obtaining access to an ethical and well-designed intervention, finishing as one of 11 semi-finalists. Now that the GLEXP has concluded, we will seek partnerships with NGOs and other organizations with a track record of improving literacy outcomes in the developing world.

AI Technologies: We have entered our core AI technologies—the digital personality—into the IBM Watson AI XPRIZE. We are currently one of 30 teams left in that competition. While primarily intended to validate our current technology, we hope the experience will enable additional academic or corporate partnerships once the competition ends. In addition, we are exploring a number of AI-related incubation programs to further develop our technologies.

Market Challenges: We are 2-time graduates of Y Combinator’s Startup School program. We continue to participate in a variety of incubation and mentorship programs with he goal of achieving product-market fit.

There are two main areas where Solve can help us.

The first is mentorship. We need to find a sustainable and profitable business model. We have brought a product to the pilot stage, but now we seek mentorship to refine and validate our fit in the market. How do we reach our customer? How do we differentiate ourselves in the crowded preschool educational technology space? How do we validate the impact of our work? How do we achieve the elusive product-market fit?

Second, once we have found the market fit, we seek the partnerships we will need in order to successfully scale our solution. For example, we are seeking partnerships with organizations with a track record of improving literacy outcomes in the developing world. We are seeking content partners to help produce content that children the world over will love. We are seeking research partners to help us keep pace in the rapidly changing world of applied Artificial Intelligence.

AI / Machine Learning Partners:

• Corporate: Microsoft Research (Machine Learning), Deepmind (Reinforcement Learning), Element AI

• Academic: MIT CSAIL, University of Montreal MILA, Stanford AI Lab, Vector Institute, CMU AI

Literacy-oriented NGOs

• Barbara Bush Foundation, CODE, Room to Read, National Center for Families Learning (NCFL)

Content Partners

• Sesame Workshop, Book Dash

One of the core components of our solution is a “Digital Personality”, consisting of a Bayesian decision engine (for learning user preferences and exhibiting device personality), and a pedagogical engine, (for making decisions balancing user engagement against ,progress through a curriculum).

To train these engines, we make use of a multi-agent simulation. A “User Agent” is trained on usage data and makes it possible to simulate vast numbers of interactions with the software. “Preference” and “Pedagogical Agents” are then trained using a variety of machine learning algorithms until the desired characteristics are learned.

While we have done our best to bring a modern set of AI technologies to bear on this problem,  the pace of change in the AI space can only be described as frenetic. We would love to be able to investigate the potential of deep reinforcement learning (DRL) to train our pedagogical agents; however, deep learning can be highly resource-intensive. If successful in winning an AI Innovations prize, we would use this prize to obtain the resources necessary (hardware, software, and potentially consulting services) to build a modern DRL training pipeline.