A non-technical guide to the tools and thinking behind Athena
I'm a parent who also happens to write software. When my 12-year-old needed help with their KS3 Biology and Maths homework, I didn't want to point them at ChatGPT. Tools like that just hand over the answers, and the student learns nothing. So I built Athena: an AI tutor that teaches through questions, not answers.
A lot of educators have asked me how I actually built it — not the code, but the thinking. What tools did I use? How do they fit together? Could something like this work in a school? This post is my attempt to explain it all without assuming any technical background. If you've ever used a smartphone, you already know enough to follow along.
You don't need to be a programmer to understand how Athena works. Think of this post as a tour of the building — I'll show you each room and what it does, without asking you to rewire the electricity.
What Athena Actually Does
Athena is a web application — a website your child opens in their browser, like Google or YouTube. But instead of showing videos or search results, it has a chat window where students can have a conversation with an AI tutor named Athena.
Here's a typical homework session:
- The student photographs their homework with their phone or tablet camera.
- Athena reads the photo and shows the student what it thinks the question says. The student confirms or corrects it.
- The tutoring conversation begins. Athena asks guiding questions instead of giving answers.
- If the student gets stuck and frustrated, Athena offers a deal: "I'll give you a hint, but you have to explain it back to me afterwards."
- After receiving a hint, the student completes a short activity to prove they understood — either reordering steps or explaining the concept in their own words.
The difficulty adjusts automatically. If the student is doing well, Athena asks harder questions. If they're struggling, it breaks things down into smaller steps. All of this happens behind the scenes.
The Building Blocks (In Plain English)
Every application is made of layers. Think of it like building a house: you need foundations, walls, plumbing, electrics, and furniture. Athena has five main layers, and I'll explain each one using everyday comparisons.
1. The Website Itself
What it is: Next.js is the framework I used to build the website. Think of it like the building plans and construction method for the house.
When you visit a website, your browser is actually talking to a computer somewhere else (called a "server") that sends back the page you see. Most websites separate the "what the user sees" part from the "what happens behind the scenes" part. Next.js lets me build both in one place.
The part you see — the chat window, the buttons, the difficulty dial — is built with a tool called React. It's the same technology behind Facebook, Instagram, and Netflix. It makes the interface feel responsive and smooth, like an app rather than a clunky old website.
The part you don't see — the server side — handles all the clever stuff: sending messages to the AI, updating the difficulty level, saving the conversation, and checking the student's work. When your child sends a message, it travels to the server, gets processed, and comes back with Athena's response. This all happens in about one to two seconds.
2. The AI Brain
What it is: Google Gemini is the AI model that powers Athena's conversations. Think of it as the tutor's brain — it understands questions, generates thoughtful responses, and can even read photographs.
You've probably heard of ChatGPT. Gemini is Google's equivalent. I chose it for three practical reasons:
- It's free for personal use. Google offers 1,500 free requests per day. For one child doing homework, that's more than enough. I don't pay anything to run Athena.
- It can read photos. When your child photographs their homework, Gemini reads the handwritten or printed text from the image. This is the same technology your phone uses for Google Lens.
- It has built-in safety filters. I've configured the strictest safety settings available. If a student tries to steer the conversation away from schoolwork, the AI politely redirects: "I can only help with school subjects. Shall we get back to Biology?"
Important for educators: Athena doesn't just chat freely. I've given the AI a detailed set of instructions (called a "system prompt") that tells it exactly how to behave at each difficulty level. At the easiest level, it gives tiny steps and fill-in-the-blank questions. At the hardest level, it asks open-ended questions that require proper reasoning. The AI follows these instructions for every single response.
3. The Curriculum Library
What it is: RAG (Retrieval-Augmented Generation) is a technique that lets the AI look up real curriculum content before answering. Think of it as giving the tutor a textbook to reference, rather than relying on memory alone.
This is possibly the most important part of Athena for educators to understand. A general AI like ChatGPT was trained on the entire internet. If a UK student asks about enzymes, it might answer using American terminology or include content from a university-level course. That's confusing and potentially misleading for a Year 7 student.
Athena solves this by having its own curriculum library. I've written curriculum content files for each topic — currently cells, body organisation, and fractions — following a consistent template:
- Key Knowledge: The core facts and concepts the student needs to know
- Common Misconceptions: Mistakes students frequently make (so Athena can watch for them)
- Key Vocabulary: Subject-specific terms with Year 7-appropriate definitions
- Worked Examples: Step-by-step solutions the AI can reference when helping
When a student asks a question, Athena searches this library first. Only if it can't find the answer there does it fall back to a general web search (limited to trusted sites like BBC Bitesize and Khan Academy). And if it still isn't sure, it says so honestly: "Double-check this one with your teacher."
Why this matters for schools: The curriculum content could be written or reviewed by teachers. Imagine a Maths department creating their own content files aligned to their scheme of work, with their preferred methods and terminology. The AI would then teach using exactly those methods. This is fundamentally different from a general chatbot.
4. The Student Record
What it is: The database stores conversations, difficulty levels, and study streaks. Think of it like the student's exercise book — a record of their work that the tutor can look back on.
Athena remembers things within a session: what the student has said, what difficulty level they're at, whether they owe any verification exercises. This is stored in a database — essentially a structured digital filing system.
I've made deliberate privacy choices:
- No personal data is collected. Athena never asks for the student's name, school, or email address. There's no login.
- Conversations are automatically deleted after 48 hours. An automated process runs every night at 3am and removes anything older than two days.
- Photographs are never saved. When a student photographs their homework, the image is processed and then discarded. It never touches the hard drive.
- Study streaks persist. The only thing that lasts is a simple record of which days the student studied each subject. This powers the motivational "streak" display (like Duolingo's daily streak) without storing anything personal.
5. The Hosting
What it is: Vercel is the service that runs Athena on the internet. Think of it as the plot of land the house sits on — it provides the electricity, water, and address so people can visit.
Every website needs to run somewhere. Vercel is a hosting service specifically designed for the kind of website I built. It handles all the server management automatically: if 10 students used Athena at once, Vercel would scale up to handle the load, and scale back down afterwards.
Vercel also runs the automated tasks, like the nightly conversation cleanup. And it's free for personal projects, which means the total running cost of Athena is zero.
How It All Fits Together
| Layer | Tool | What It Does (In Plain English) |
|---|---|---|
| Website | Next.js + React | Builds the pages and chat interface the student sees, and handles all the behind-the-scenes logic |
| AI Brain | Google Gemini | Understands questions, generates Socratic responses, reads homework photos, follows safety rules |
| Curriculum | RAG + File Search | Gives the AI a library of real KS3 content to reference, so answers match the actual syllabus |
| Student Record | SQLite + Prisma | Saves conversations and difficulty levels during a session, auto-deletes after 48 hours for privacy |
| Hosting | Vercel | Runs the website on the internet, handles scaling and automated tasks, costs nothing for personal use |
| Animations | framer-motion | Makes the difficulty dial spin smoothly, notifications slide in, and progress rings fill up |
| Maths Display | KaTeX | Renders fractions, equations, and mathematical notation properly (not as ugly plain text) |
What This Means for Education
The technology behind Athena isn't experimental or expensive. Every component I've described is freely available. The total cost of building and running Athena is zero pounds per month. The skills required are significant — this took real software development knowledge — but the barriers are financial, not technological.
A few things I think are worth highlighting for educators:
-
The AI can be constrained. This isn't a free-for-all chatbot. I gave Athena detailed behavioural rules: use Socratic questioning, never give direct answers without verification, detect frustration and offer structured help. The AI follows these rules consistently. A teacher could define their own pedagogical approach and have the AI follow it.
-
The curriculum content is teacher-controllable. The RAG library means the AI teaches from content you provide. If your Maths department uses a specific method for adding fractions, you write that into the content file and the AI uses that method. This is not "AI knows best." This is "AI teaches what you tell it to teach."
-
Privacy can be designed in from the start. Athena collects no personal data, requires no login, and automatically deletes conversations. These aren't afterthoughts — they're foundational design decisions. Any educational AI tool should be built this way, especially for children.
-
The hard part is the pedagogy, not the technology. Building the chat window took a few hours. Designing the negotiation system — how Athena detects frustration, offers hint trades, requires verification, and adjusts difficulty — took weeks of thinking and testing. The educational design is where the real work lives.
Frequently Asked Questions
Could this work for a whole class?
In its current form, no — it's designed for one student with no login system. But the architecture could be adapted. You'd need user accounts, a proper database, and a paid API tier (Gemini's free tier handles about 1,500 conversations per day, which would cover a class but not a whole school).
How much does it cost to run?
Currently, nothing. Google's free API tier and Vercel's free hosting mean the monthly running cost is £0. For a whole school, you'd be looking at API costs based on usage — probably tens of pounds per month, not hundreds.
Can it teach any subject?
Right now it covers KS3 Biology and Maths because those are what my child studies. Adding a new subject means writing curriculum content files and creating a new AI content store. The tutoring logic (Socratic questioning, difficulty adaptation, verification) works subject-independently.
Is the AI always accurate?
No AI is perfect. That's exactly why I built the three-layer knowledge system. When Athena answers from the curriculum library, accuracy is high because the content was written by a human. When it falls back to general knowledge, it flags this honestly. I would never suggest replacing teacher oversight — but as a homework companion, it's significantly more structured than a student Googling the answer.
Could a school build something like this?
With a developer, yes. The tools are all free and well-documented. The harder question is who writes the curriculum content and who designs the pedagogical rules. That's where teachers are irreplaceable. The technology is the easy part.
What's Next
This is one post in a series about building Athena. The next one dives into the negotiation engine — the system that detects when a student is frustrated and offers them a structured deal instead of just giving up or handing over the answer. It's the feature I'm most proud of, and the one that required the most thought about how children actually learn.
If you're an educator interested in this kind of thing, I'd love to hear from you. The technology is ready. The question is what we do with it.
If this was useful, share it with a colleague. The more of us who understand what's possible, the better we'll build it.
