Edsy AI English Coach

English fluency has become a status symbol in Thailand/Asia because it is key to economic and social mobility. People who speak good English get better jobs, income, knowledge, skills, networks, opportunities, etc. 

However, the majority of 10 million Thai students cannot communicate in English (only <2% enrolling in international schools can) because traditional English classrooms are ineffective. Most teachers cannot speak English so they lecture in the local language and focus on grammar. Students do not have the opportunities to individually practice speaking, get feedback, and learn from trials and errors, which are critical for any skill development. 

Hence, Thailand is consistently ranked among the worst, at 101st out of 113 countries according to the EF English Proficiency Index in 2023, far behind India at 60th, Indonesia at 79th, China at 82nd, etc. 

With the world becoming more globally connected than ever + the Future of Work being powered by remote/hybrid work and cross-border sharing economy/freelancers, English language will also become even more critical for everyone from any country.

In our approach to enhancing traditional English classrooms, students engage in active learning by first listening to lectures per usual (learning vocabulary and grammar concepts), and then practice speaking individually with our AI Coach. The content is synchronised with their classroom lessons, allowing them to apply what they've learned in real practice.

The AI Coach provides open-ended prompts (not just reading-out-loud exercises), offers scaffolding when necessary, and gives immediate feedback on their responses. It complements regular English classes, acts as additional "Communicative English" classes, or is used for after-class assignments. It can be used on any device with minimal internet speed. 

For example, if students learn about "jobs" in their English class, they might listen to lectures on related vocabulary and grammar but lack opportunities to speak individually. Our app fills this gap by providing related prompts like "Tell me about your dream job", which students self-record responses to. If their initial response doesn't meet the AI Coach's bar for fluency and relevance, it offers appropriate scaffolding like simplifying questions, teaching essential vocabulary/sentence structures, or providing full/partial model answers until the response improves. Afterwards, the AI Coach gives feedback on grammar, suggestions for sounding more natural/native, and pronunciation. Once the student completes this cycle, they move on to the next prompt.

After the lesson, students can review what they've learned, track their points, and engage with gamification features.

To come up with this AI Coach, we developed (1) Machine Learning classifiers to evaluate learner’s verbal responses and (2) built an Intelligent Tutoring System leveraging GenAI to provide targeted scaffolding and feedback. The system also uses the learner's history and knowledge tracing to suggest personalised practices based on spaced repetition and Knowledge-Learning-Instruction framework (KLI) from our team's PhD research at Carnegie Mellon University: https://scholar.google.com/citations?user=a9l8VDAAAAAJ&hl=en

5 million Thai students between grade 4-12 in public schools and 500,000 who are out of schools (mostly those in/near poverty). Our solution provides students with individual speaking practices (answering open-ended questions, not just reading out loud from scripts) and necessary scaffolding + immediate feedback so that they don’t just “know” but can really “use” the language. 

However, our solution is also now being used in public/private universities as well as being used by relevant government entities (e.g., Bangkok Metropolitan Administration) to train the school teachers too.

We plan to expand to other Asian markets by the end of this year to serve many millions more. 

All of us attended public schools and had first-hand experience with the ineffectiveness of traditional English class. However, we were fortunate to have gotten scholarships to study overseas and eventually picked up English. We have also been volunteer teachers, including at rural schools with limited access to electricity, so we understand the issue deeply from both the learners’ and teachers’ perspectives. One of us has deep expertise in learning science and AI in Education from his PhD research at Carnegie Mellon University. Human-Computer Interaction Institute 

Furthermore, we have been working closely with the Bureau of Educational Innovation Development, Ministry of Education, to conduct an R&D project with 12 public schools, collecting insights/feedback from over a thousand students, teachers, and schools’ management to improve our AI solution. The joint report will be published in May and presented to the Ministry’s senior management. So we ensured that the solution is meaningfully guided by the communities’ input, ideas, and agendas, to the maximum extent possible.

More than 30,000 students from 70+ schools/institutions have used our AI solution (full/partial). This spans public/private schools and also a few higher education institutions in need. This includes schools and institutions under the Ministry of Education, the Bangkok Metropolitan Administration (BMA), Thailand Education Sandbox (Rayong Province), Connext ED Foundation (a national PPP initiative), and AFS Intercultural Studies Foundation.   

We are also working with the Bangkok Metropolitan Administration and the Equitable Education Fund (EEF), the country’s main body responsible for reducing educational inequality and developing teachers’ effectiveness, to further scale up towards 300,000 students by the end of this year.

Getting support on scaling outside of Thailand to other Asian (and beyond) countries through MIT-Solve’s vast network of policy makers, civic society organisations, impact investors, and advocates. 

Access to a network of impact-driven educators and change-makers to share knowledge and collaboration opportunities, especially on learners’ engagement, inclusive pedagogy/learning design, impact measurements, etc. 

Gain knowledge and skills particularly related to public advocacy and government/stakeholder relations.

Open-ended prompts, not script-reading:
Most existing solutions focus primarily on reading aloud from scripts, which aids in pronunciation but doesn't necessarily prepare students for engaging in English conversations. Through our R&D efforts with the Ministry of Education, we discovered that solely focusing on script reading and pronunciation practice is insufficient. The critical aspect of language mastery involves not just pronouncing words correctly but also constructing sentences, selecting appropriate vocabulary and phrases, and negotiating meaning in conversations. The biggest barrier to fluency is often not pronunciation, but coming up with responses in sentences. 

Scaffolding:
To maximize learning outcomes, each learner must get relevant scaffolding activities (e.g., simplifying questions, full/partial model answers, vocab building) that fill their specific gaps, rather than getting all the same patterns regardless of individual learning needs. That’s why we developed our own ML classifiers that can evaluate responses and pick suitable scaffolding. Most players only rely on third-party LLM’s wrap/API and don’t provide much scaffolding that is critical for language acquisition.

Modularity:
100% self-learning with AI does not work for everyone, especially young learners who are used to interactive/engaging instructions. That’s why we model our AI Coach to be modular such that it can easily sync with the school’s/university’s English class and/or engaging TikTok-style video clips, thereby maintaining human-like elements that make learning authentic and fun.

Our solution has been developed through R&D efforts with the think tank unit of the Ministry of Education and has undergone pilot testing through the Education Sandbox initiative, a government-led effort aimed at revolutionising the education system. With this solid foundation, our solution has the potential to be implemented nationwide, transforming English language education in Thailand and beyond.

We recognise that teachers may be resistant to drastic changes in their teaching methods, especially if they have low proficiency in English. However, they excel in teaching vocabulary and grammar. Our approach introduces the AI English Coach as a supportive tool that seamlessly integrates into existing teaching practices, minimising disruption and resistance. Students continue to learn in their usual manner but are provided with opportunities for individual speaking practice and feedback through the AI English Coach, which acts as a teacher assistant both in and out of the classroom. 

Our theory of change centers on the AI Coach addressing a critical gap in traditional English classrooms by providing students with opportunities for individual speaking practice and feedback. As students engage in regular practice responding to open-ended prompts, they develop increased confidence, accuracy, and fluency in English communication skills, as measured by CEFR levels. Data from our pilot phase in 4Q23-1Q24 showed an average improvement of approximately 15% in fluency and accuracy within a short period, validating the effectiveness of our approach.

With enhanced English communication skills, students gain access to better job opportunities, higher incomes, expanded knowledge, skills, networks, and broader opportunities overall. This assertion is supported by existing research from reputable sources like Education First, Pearson, and Harvard Business Review. Additionally, we plan to conduct further research in collaboration with relevant authorities like Thailand's Equitable Education Fund (EEF) and Thailand Development Research Institute (TDRI - Thailand's foremost policy think tank) to strengthen the connection between our students' improved English communication skills and their increased income and job placement rates.

Furthermore, the AI Coach has been successfully utilised to train English teachers in Thailand (e.g., through projects with Bangkok Metropolitan Administration and AFS Intercultural Foundation). This initiative holds even greater potential impact, as proficient English-speaking teachers can deliver lectures in English, shifting the focus of the classroom from traditional grammar-translation approach to communication skills. This not only enhances the learning experience but also prepares students more effectively for real-world communication scenarios.

We aim to improve English language proficiency levels, as outlined by the Common European Framework of Reference for Languages (CEFR), among primary and secondary school students, college students, and teachers. Our goal is to achieve the Ministry's guideline across all these groups e.g., B1 for high school students and B2 for teachers. Currently, >80% of all students fall short of the guideline and 75% of English teachers are still at A2, according to British Council. 

To measure our progress, we developed a formative assessment tool similar to the TOEFL iBT exam but with minimal cost, powered by ML classifiers that serve as an "AI Grader" (rather than using certified human-graders that cost more). This tool serves as both a pre- and post-test to evaluate improvements in language skills.

Improvements in CEFR levels serve as tangible indicators of learners' mastery of the English language and communication skills, opening doors to broader knowledge and opportunities. Our impact goals align closely with Sustainable Development Goals 4 (Quality Education), 8 (Decent Work and Economic Growth), and 10 (Reduced Inequalities). We seek to measure the link between our students' improved English communication skills and higher income/job placement to underscore the ultimate impact. 

Intelligent Tutoring System + ML classifiers that include an Automated Speech Recognition (pre-built), Feature Extractor (own; for both acoustic and language features), Scoring Model (own; ensemble method), Grammar Checking Model (own), LLMs (open-sourced and further trained through research collaborations with local academia, to ensure lowest cost structure possible). Wraps of best-in-class pronunciation (for maximum accuracy) and text-to-speech (for maximum authenticity) models.

12 full-time staffs including the CEO, COO, CTO/AI Lead, 3 software engineers, 1 product management, 3 business development, and 2 business operations.  

4 part-time contractors include 2 engineers and 2 freelance learning designers.

About 1 year since we did a full pivot from the original online tutoring platform and started working on AI-based solutions. 

From the ML classifiers we had, we quickly turned it into the AI Coach MVP in 2 months between September-October 2023, launched in November through a free pilot with 50 schools totaling >3,000 students. First conversion - despite being an MVP -  in January 2024. First full product release in May 2024. 

3 co-founders are 2 male and 1 female. Of the 14 full-time team members, 6 are female, 1 is LGBTQ+. 

We offer our AI Coach directly to schools, institutions, and relevant government entities, with pricing ranging from $2 to $30 per student per academic year. The cost depends on the specific needs and use cases of the institution, such as desired features and the number of practice sessions.

Our pricing structure provides flexibility to accommodate varying budgets and requirements. Despite our competitive pricing, we maintain significant cost advantages over other providers. This advantage stems from our in-house development of multiple AI engines, such as the ML classifiers and grammar-checking models. Unlike many competitors that rely heavily on third-party APIs, our independent development enables us to offer an affordable and accessible solution without compromising on quality or functionality.

We've achieved financial sustainability and are currently operating at break-even, given our relatively small team/overhead costs and big streams of revenue from major institution accounts.