Introduction

Hi, I'm Abishan.

I'm here with Ibrahim, my friend and co -founder of our business called Compass Maths.

Entering the Education Space with AI

So, we just decided to go into the education space and try and really bring AI to the education space.

That's kind of slowing.

It hasn't really made a big step into it yet, and that's what we're trying to address.

So, next slide, please.

The Problem with Current Homework Tools

So, what we built and a bit of context about why we've built it.

So starting with the context, currently in secondary school education, there seems to be a bit of an issue where homework tools aren't really making the most of the AI technologies that are coming forward, and they're a bit outdated.

So as a tutor right now, a lot of the feedback I get from students is that their homework tools feel a little pointless.

So they receive a question, they write the answer, if they get it right, well done.

If they get it wrong, a mark skin is given to them, and they don't really receive any personalized feedback.

So they don't know where their working hours went wrong.

They don't really know what else to expect rather than just what the mark seems looking for

So that kind of creates like a bit of a stagnation in their learning and that's the problem

We're trying to solve.

Our Solution: Compass Maths

How It Works for Students

So what we built was a system where teachers are able to

Just use questions that from databases that are provided online and set this as homework

the bit that's different for us is that we allow students to enter the written working out for each question and

that is fed into an API which

which comments on their working out and provides direct feedback onto what they're doing wrong

rather than what the marketing team is expecting you to do.

So that creates this little personalised feedback loop which the students are looking at right

now.

So in that little box over there is basically what the students see.

How It Helps Teachers

And building off this, what we also have is a teacher section where they're able to see

how a class is performing.

So right now some teachers would just mark the homework and continue with a syllabus

of what the school expects them to teach.

And to an extent that's good where there's a syllabus

that they need to be following.

However, none of the lessons are really like personalized

to a scale.

However, we've created the AI to basically ask,

you get prompted to ask, how is this class struggling

and what are any specific topics I need to teach?

And we're able to generate lesson plans

for those classes in particular.

This makes the lessons more like interactive

interactive, and much more personalized to what's actually happening with your class,

and that is shown to improve learning.

Next slide.

Yeah.

I'll just do it from here.

Challenges and Safeguards

Protecting Student Data and Privacy

Yeah, so some of the issues that we faced, so when you're working with student data,

you have to be really careful about user data and everything.

Obviously we don't want our kids, well, we don't have kids, but your kids' information

information, you don't want your kid's information just lurking around, so essentially one industry

standard is this long word called pseudonymisation, I don't know if I, basically it just means

like when we feed, so when the AI tools are queried for student data, it's not, the student's

information won't be fed to the model, so for example one of our tools is we have a

chatbot where teachers can go in and ask about how their students are doing specifically um so you

don't want to go in there and be like oh how's emily doing at this specific subject that's why

if you uh if you see your kids um the tools your kids use at school um they usually create a user

with like a name like i don't know like like i don't know um rainbow zebra or something like that

or like i don't know sparkle angel and like that's their username that's essentially to protect them

from having their information online so that's something we did and then another uh another

Accuracy, Hallucinations, and Validation

Another technical challenge we had, the second and third, hallucinations and validations,

those kind of are under the same bracket where we need to make sure that the information

we're getting from the user, so basically they're handwritten working out, and the information

we're giving them back, so their feedback is correct.

One way we've done this to protect from hallucinations is we've set up basically a rag model system,

them so it's not like we're taking the students data yeah the students data and they're working

out and then we're not and it's not coming straight from a model we basically have the

specifications so AQA edXL basically the entire mark schemes and then the agents then go in

and then be like okay this is the question I have there's a vector database it's basically

goes in and it finds oh this is the specific part of the mark scheme that this child's working on

on, brings it back, and then the LLM is working from the user's answer, the actual answer,

and then maybe in the future we get somewhere where the LLMs are so smart that you don't

even need rag models and they do everything, but I think we're quite away from that, so

for now this is one way we're protecting ourselves from that.

We also have a page where we confirm the students' marking out, so we literally ask them, is

Is this what you wrote?

And that's actually something that LLMs are really good at,

is the OCR, the image, like reading an image

and then printing out, yeah, this is what they put.

So we just had those in the way just to make sure

that we're really hitting those 98%, 99 % thresholds

in terms of accuracy.

Yeah, this is actually the first one.

Pseudonymisation in Practice

Yeah, so going back to the data privacy aspect,

so this is a screenshot from our app

that we're going to be demoing later.

So just going back to it.

Over here, Blue Penguin, Silver Fork, Scarlet Phoenix, and Quick Rabbit are all just names

that are generated, and we assign it to students.

The main thing is that these names are never, ever fed into the LLMs that we're using.

This keeps it completely private.

Only the teachers and the external database are the only places where we can really access

this.

This prevents us from ever having a leak, or ever data getting leaked of students' names

games and what schools they go to, and that is something that we just completely want

to avoid.

That is kind of why we went for this approach.

We will see more of this in the demo, and the next one.

RAG Pipeline and Model Guardrails

Just continuing on, it is never the user to the AI model straight connection.

We have the specification there.

From testing, we have been getting really good results from this model, the rag model,

model.

So that's just one way we're protecting ourselves.

Handwriting to Math: OCR and LaTeX Confirmation

And yeah, this was the page I was

talking about.

So once we get the users working out, we basically convert it into math notation

called LaTeX, which is basically something that helps computers really understand mathematical

notation.

And then we convert that back into normal math notation, which is what you see

on the right.

So the users from here can be like, yep, this is what I put down, just so

we don't want them to get into the final slide and be like, oh, this feedback's incorrect

correct because they didn't understand

what I said.

So then we have an option where it said, no, let me fix

that, where we have

a maths keyboard where you can go in and basically change

it to, yeah, this is what I meant instead.

Yeah, I think

it's time for the demo.

Hopefully it works.

If it doesn't work, we have a recording

of it working perfectly.

It was working

like 30 minutes ago, so it's not like

this recording was from like three weeks ago.

All right, here goes nothing.

Live Demo Walkthrough

All right, so

So basically, oh, I forgot about that.

Selecting Questions and Submitting Work

Yeah, so the user can select a question.

Right now, we only have three questions we're testing with,

but we can get to a point where we have, like,

all the different topics.

But yeah, it works for one, so it should work for all.

So then the user goes in.

They upload their work, you know.

Then we patiently wait for the people who,

for the API response.

Perfect, it doesn't take that long.

Diagnosing Mistakes and Delivering Feedback

Yeah, so we get to this page here.

What we found is this page is quite decent

in terms of accuracy.

We're getting very accurate results.

So basically the students,

so just to give you guys context,

this answer is correct, is incorrect.

And I mean, does anyone wanna like try guess why,

or like knows why it's incorrect?

We'll give you like a prize.

We'll give you like a bunch of free Diet Cokes.

works.

Yeah, so from our GCSE days, the line that says 1 equals minus 9 plus C, you're

meant to move the 9 over, which becomes positive, so it's 10 equals C.

But yeah, that's why

this tool, from the teacher aspect, is really positive in terms of saving teachers' time

marking no one wants to go in just be marking messy handwriting and like after a while it's

human fatigue your teachers are naturally just going to look at the last line be like did they

get it right did they get it wrong tick move on move on this really makes it so that we're

working in as much video as possible okay so is this correct yes yes correct so this is a

This is the important bit.

I hope this goes well.

Okay.

Oh, yeah.

It did get it.

So it gave them the right answer,

y equals minus 3x plus 10,

and then it gives you the feedback of what you did well,

and then it gives you the areas to improve.

For this specific question,

I think for different...

What are they called?

But exam boards, they give you different marks.

So I think this is the Edexcel exam board,

which gives you one out of three,

and then I think AQA and the others

give you two out of three marks,

if this is what you put, which is really harsh.

So what we're doing is, when we're speaking to schools,

we're asking them what exam boards you're working with.

Luckily, when it comes to maths,

it's pretty common to be using Edexcel.

A lot of the schools use Edexcel,

so we'll basically just work with

what the main exam board is.

But we're always ready to provide personalized solutions for schools.

Verifying Correct Solutions

So then we're going to go in here.

We're going to go back to that exact same question.

And this time, where did I save it?

Okay, here we go.

I've now uploaded the correct answer.

And we're going to just go through this process again.

again.

Yeah, so similar to last time, you just check if, so the student will just check

if everything they wrote, yep, yep, yep, yep, yep.

Okay, so that's lining up.

We should

get a correct answer.

Please.

Oh.

Okay, well, this is not good.

I'll try that one more time,

if not, then.

Okay.

Yeah.

We probably have to, like, probably have to go in and make

sure that doesn't happen again.

This is the correct answer.

The student got full marks.

Areas to improve.

Usually, when you get full marks, it doesn't give you any areas to improve,

but the models are so accurate that sometimes they'll give you even a little response in

terms of, oh, you could have factorised this more in terms of the examples.

Sometimes they're

kind of harsh with stuff like that.

But yeah, so the answer's correct.

And yeah, essentially,

Essentially, this is the student's perspective, and this is the tool that the students use.

We've gotten to a point where it's pretty accurate.

If we do this, I don't know, like 20 times, I think 18 out of 20 times, it will be perfect.

Sometimes we have to go in and there's API errors and stuff like that, which are usually

not the model's problem.

Because of the RAG model, because of the authentication and the confirmation, we're getting really

good results from that.

that.

Teacher Dashboard and Insights

And then the teachers aspect.

So this is basically something that the teachers will

see.

So right now, we're using sample data from like, so examples release, like common

mistakes students make every year.

So some of the stuff we have, we have right now.

So

teachers can track like the class averages for specific subjects or for all homework

work as a whole and then they can really see like this is where this is where i think are the tools

really are useful because the teachers get really specific understanding of where the class is

struggling with so the teachers can see here like the different the different topics um and then so

here in the common mistakes you see stuff like sign errors in algebra or like factor um fraction

operations or like uh using wrong formulas for areas such volume so uh yeah so when you have

questions in gcc for like i don't know find the volume of a cylinder find the volume of a cube

we'll find the volume of all these things there's some there's there's quite a few formulas um

so like we're getting really really specific responses in terms of what the students are

struggling with it's not like oh the class isn't good at trigonometry the class is not very good

at i don't know vectors like we're getting really specific and um so we've gone a step further

Chatbot for Class and Student Support

further, and so we have a chatbot, we have a chatbot, so this is like, we implemented

this tool like a week ago, and it's, right now, we don't have it live, so I don't think

I can demo it, I don't even know if I have a video of it, I probably should have, but

so basically, so we asked why is Quick Rabbit struggling again this week, I noticed the

score dropped specifically, so this is the whole thing about protecting user data, we're

We're not going to be going into the LLMs asking,

putting sensitive information in there.

And then the LLM responds with something like,

based on this person's submissions,

they're struggling with multi -step trigonometry equations,

particularly with combining Sokoto with Pythagoras theorem.

So those are like, I don't know, I did it like six years ago.

So you can see specific key issues identified,

confusion, stuff like that.

You can track how they're doing compared to the rest of their class,

and then it gives you some recommended actions,

Generating Targeted Lesson Plans

or you can just go in and generate a lesson plan.

which will basically generate lesson plans

for, like, specific topics that students struggle with.

Yeah, and you can also track...

Performance Tracking

We have a student performance thing,

which this is just, I don't know,

this is just something that I think major,

like, all the other competitors provide.

So that's what we have today.

Closing and Q&A

We will field some questions.

But, yeah, this is our LinkedIn.

We're still at a point where we're talking to people

who, if you're in a position where you have ideas

is or you have people who are interested in this kind of sector then please let us know uh we'll

Market Landscape and Differentiation

try to contact everyone um but in terms of just before we get to questions one thing we will ask

we'll let you know is uh in terms of competitors no one is doing what we're doing right now in

in terms of maths uh there are other tools out there for stuff like biology and stuff

and in general um they're usually business to customer tools so going straight to students

and ours is more of a business to business we're going to go to schools and get them to use that

in terms of traditional competitors in the school market usually what you have

is like you have tools like dr.

Frost and Hegarty maths where it's just like a

it's a screen has a question has a box you put the answer in tells you yes or

no you continue ours is much more detailed in terms of your game specific

responses on your working out and then there's obviously the teacher aspect

yeah that's it for now yeah when I feel some questions if there are any