FundaAI - An AI School, building the first AI Tutor

I’m building an AI School for self driven kids across Africa. The future of education is kids having access to personal AI Tutors designed to help them improve their grades, build critical thinking skills and learn foundational technical skills for an increasingly technologically driven world. A lot of EdTech solutions have failed to scale continentally because the underlying problem has always been the same, EdTech traditionally does not address the lack of consistent, fast and cost effective internet access. An “AI School” of laptops with an AI brain interfacing with AI Tutors leap frogs these barriers, providing customized education to students.

I am starting off by giving my nephew a laptop with an offline AI as it’s “brain” and an App Store to give him access to a “virtual campus” of AI Tutors (desktop apps).

For context here’s a breakdown of what I have been building

1. What is FundaAI

2. Building the “Virtual Campus”

The First AI Tutor; The Examiner

The Examiner is an AI Tutor to help students improve their grades through practicing past paper questions. The tutor, generates exam questions, using over 2000 past papers as its “repository of information”, it maps the questions to answer sheets for each exam paper, enabling students to answer random past paper questions and immediately receive feedback from AI marking the student’s efforts.

The marking initially occurs offline using the local AI model, but the output is queued so that whenever the user is online, the output is sent to a larger model for a more detailed user report. This is an electron desktop app. This requires its own blog post to explain how it works. For now, here’s a high level of summary of the business questions guiding the architecture of this AI Tutor;

• I need this to work primarily offline to reduce the need for a constant internet connection - this means building a local cache of questions and answers. There is a SQLite database to track metadata (we only want to refresh the cache after the user has answered 75% of the questions. This means we need to track the questions the user has completed), the answers and questions are stored locally as JSONs and the assets (images, graphs, tables and charts are downloaded and stored separately). We use the SQLite database to map the images to their respective questions.

• To practice for exams you need to answer different variation of questions - this means syncing local content with content from a larger repository of information (stored in MongoDB) and tracking what the user has completed to avoid repetition.

• The AI Tutor must be able to give the student quick and accurate feedback - this means a dual process where initial gradings and feedback are fed back to user through a local/offline AI Model. Since the first pilot laptop is a 2013 MacBook Air 8GB RAM (it has to be affordable to the end user), I found the best model to use was 1.5B 4-byte quantized Deepseek model.

The Dagger AI Agent - Extracting individual questions and answers from past papers

Dagger recently released support for building AI Agents through Dagger. I see my Agent as more of a pipeline requiring;

1. reproducibility: I need to run this whenever add more past papers to my AI Tutors “memory base”.

2. seamless chaining of a pipeline of steps: I need to run steps sequentially, to discover new papers, download them and extract the text and images from each paper. I need to ensure that the questions are correctly grouped (include all sub questions, images, graphs and tables associated with the questions)

3. caching each step will ensure that failures don’t require me to restart the entire pipeline allowing for quicker processing

For these reasons I opted to give Dagger a shot at building an agent to discover and categorize exam question papers and answer sheets from the Cambridge International curriculum, covering all subjects and levels and extracting and correctly grouping all individual questions associated with an exam paper.

This is done and mapped to the answers from the answer sheet relevant to the question paper. Through this I build a dataset of questions and answers to feed into The Examiner AI Tutor

Current Progress

Here’s a short video of what I have built with the Examiner. Currently, I am working on prompt engineering and or fine-tuning the local model to get more accurate responses that can conceivably be used to produce a preliminary report on a user’s performance on a question paper.

These are the last steps before I ship the laptop to my nephew in Zimbabwe. Once this happens it becomes a question of iterative improvements based on user feedback while focusing on getting more sales to pilot to more people.

FundaAI HomePage

GitHub