Overview
In the article below, I outline how I built a series of AI-powered tutoring prototypes for South Africa’s largest online high school. The tools used are listed immediately below, while the rest of the article outlines the context and outcomes of this project.
Tools used:
- AWS:
- S3 bucket for front-end deployment
- Elastic Beanstalk for backend server
- AWS Polly for text-to-speech functionality
- Secrets Manager
- Gitlab & Gitlab CI/CD for remote code repo and deployment
- ChatGPT’s API platform for LLM functionality
- React.js for the front-end chat messaging interface
- FastAPI for the backend server
How AI can help cost-effectively overcome years of learning gaps
In South Africa, researchers estimate that the average learner in 80% of schools is approximately 4 grade levels behind by the time they reach grade 9. Put differently, imagine a learner reaching grade 9, but with the mathematics scores of a grade 5.
There is a known way to address this – in theory
As Benjamin Bloom demonstrated in 1984, small group tutoring (two to three learners at a time) produces dramatically faster learning gains than conventional large group classroom instruction. In theory, intensive small group tutoring would enable most learners to catch up their learning gaps.
The problem, which is as present today as it was then, is cost. Although differing estimates have been calculated, providing this kind of instruction to all learners is likely not far of an order of magnitude more expensive than available education budgets. This problem holds across most, if not all countries, because cost of teacher/tutor salaries tends to rise proportionally to a country’s wealth and level of economic development.
AI provides a unique opportunity
Artificial Intelligence, specifically relatively inexpensive Large Language Models (LLMs), have the potential to change this equation. If appropriately constrained and refined, digital tutors powered by LLMs may be able to provide highly personalised, always available, low-cost tutoring to every learner in a classroom.
With this insight in mind, in Q2 and Q3 of 2023, while working for South Africa’s largest online high school, I built a series of prototype AI-powered tutors which served to:
- Demonstrate how an AI-powered tutor could be used in the online high school context;
- Early lessons about how best to integrate and utilise an AI-powered tutor
- Enable prototype testing, both internal team testing and limited user testing;
- Provide initial cost estimates and functional requirements
Project Sizo
The AI-powered tutoring prototype project was given the internal name “Sizo”, which is adapted from the isiZulu word for assistance (usizo). As a side note for any isiZulu speakers, I picked this name because I wanted a two syllable name, I thought that ‘siza‘ would be easily misheard as ‘Caesar’ and I liked the double meaning that I spotted in “Ngizocela uSizo” (I will ask for assistance / I will ask for Sizo).
Three principal versions of Sizo were built. All three versions were prompt constrained to function as tutors, not providing learners with direct answers and instead employing a variation of the Socratic method (end-directed questions to help guide the learners to think of the answers themselves, with only the minimum nudging and assistance).
Full-screen tutor
This version provided a full-screen chat application, demonstrating what a standalone tutoring service could look like. This version was seen as least likely to be implemented for learners, because it would take learners away from their lesson content (a distractor). For that reason, this version of Sizo was built into a general demonstrator of a range of potential functionalities for different users and use cases. This included:
- Text-to-speech functionality synchronised to real-time text generation (to help improve learner literacy by pairing visual and auditory engagement with individual words)
- Teacher-focused use cases, such as administrative tasks and question generation
- Testing OCR (i.e. converting handwritten text to digital text) services such as Amazon Textract for AI-assisted marking exploration
- Testing different constraining prompts
Pop-up tutor widget
Integrated directly into the learning platform, this saw Sizo available as a pop-up widget, located in the bottom right hand size of the screen.
This was seen as a potentially viable use case for a tutoring functionality, but with the following caveat: The tutor would rely on the learner to initiate the interaction and would only be helpful if the tutor had adequate context on the learner’s current lesson focus to help constrain its responses. In addition, experience has shown us that learners who most need help are less likely to ask for help (although they are more likely to ask an automated service than a human tutor), which is also a downside of this version of Sizo.
Building the necessary technical services to enable this functionality is not trivial and for this reason, the pop-up version of Sizo was not seen as a viable short or medium term priority.
Sizotjie – embedded Sizo
This version of Sizo sees a highly tuned/prompt-constrained version of Sizo embedded directly into individual lessons. Each embedding contains its own unique prompt, tailored for that particularly lesson and activity. The prompt includes a specific starting question, which functions as a concept-checking question that proactively engages the learner.
This could be, for example, asking the learner for the advantages of a mixed economy. Based on the learner’s response, the tutor is able to tailor its subsequent engagements, guiding the learner towards demonstrating understanding of the concept outlined in the prompt for that particular lesson. If the learner clearly understands the concept, then the tutor responds with higher level questions to double check the learner’s understanding. If the learner’s response indicates a lower level of understanding, then the tutor responds with simpler questions and helps guide the learner to a basic level of understanding in a defined number of responses.
Another usage of embedded Sizo is to help learners practice long-form writing skills, as shown in the screenshot below. Here, learners are preparing to write a marked assessment involving a technical design brief. Embedded Sizo is able to walk each learner through the required steps in a Socratic manner, asking questions and providing reflective question-based feedback to help prompt learners to think through and discover the correct answer, without simply giving away the answer.
Conclusion
Project Sizo was handed over to the Integrations team in Q4 of 2023, to be formally included in the Product Roadmap. A strategic decision was taken to deprioritise AI-powered tutoring for 2024, due to commercial constraints (the online high school in question prioritises affordability and access, rather than more expensive, less accessible offerings). The product line is under consideration as part of the 2025 roadmap.
I remain firmly convinced of the potential of AI-powered tutoring to supercharge learning, particularly where significant learning gaps are present.