Edexia Research Project Overview

Hey, I'm Daniel, the CEO and co founder of Adexia.

And this is just a quick video sort of summarizing all of the information that I'll write up below this.

But just to give you a bit of, an overview of Adexia.

Our core mission is to try and do whatever we can to fix the education system.

And we envision a world where every student has access to truly personalized and world class education that fosters a sense of curiosity, and critical thinking and purpose rather than teaching through more obligation based means.

We think the best starting point for this is building a teacher assistant for grading and feedback to try and deburden some of the workloads that teachers face while simultaneously improving the accuracy and consistency of their assessment and enabling them to give more frequent feedback for student development.

We've been fortunate enough to work with some amazing schools around Australia, and we're now also working with a few internationally in the US and even one in Switzerland.

And we've had the pleasure of also going through the Y Combinator program in San Francisco, which also helped us raise a significant amount of funding towards the end of that program from some truly amazing sort of mentors and investors in this sort of research project.

We what we'll be looking for from the potential interns is some support on one of the biggest technical challenges that we're currently facing.

Where AI is at the point that with good enough workflows and optimization, it's effectively able to do almost any task.

But the bigger nuance is actually defining what task to do in the first place.

In assessment, there often be very nuanced and slightly vague marking guides or grading rubrics, and they'll be present with sort of terms like discerning, evaluation, and various others.

And the key challenge is actually understanding what does that actually mean in this specific context, because every different teacher will interpret it slightly differently, based on their own experience, their own classroom, and also their own biases.

And so AI will have a different interpretation to any given teacher, just like how two different teachers would have a different interpretation.

And for the technology to really, truly provide a game changing amount of value, we need to somehow bridge this gap and define this contextual information that exists in this teacher's mind.

But this is actually quite hard to define.

When you ask a teacher what does it mean for it to be a discerning evaluation?

It's sometimes hard to articulate.

Humans, are much better at knowing it when we see it, rather than precisely explaining and defining it.

And trying to get a teacher to define that can sometimes take hours and hours.

And so we need to develop a solution which can facilitate the passive extraction of this information in as streamlined, as a way possible.

And to do this, we first need to figure out what does a fully defined human thought process really look like.

What does it look like to fully define these slightly ambiguous terms like discerning evaluation in a way that doesn't over narrow the student's potential response?

Because there's different ways that we can define the rubric, we can either break down and further actually define those terms in itself.

Maybe we go from discerning evaluation to more clearly saying that means identifying and explaining three key sources.

But the issue with any kind of breaking down of a criteria is that you can sometimes over narrow the student response and prevent the curiosity of students from taking different approaches, to actually coming up with a discerning evaluation.

So we need to be very careful with trying to overly define these criteria.

And then the second thing that we can do is actually store examples of say, discerning evaluations which give context of what a discerning evaluation could be without very concretely saying it has to be X, Y or Z.

But there's a nuanced balance between these two factors and the actual balance between these can be different, based on different assessment types, as well.

And so effectively you need to figure out what exactly does that look like?

What is the ideal definition of these rubrics look like in these different cases?

And then the second step is we need to figure out what is the fastest way that we can actually define that, how do we extract that out from a teacher in the quickest, and most streamlined way possible?

And so this is what you'll be exploring in this role where you'll be actually manually going through the process of trying to define your own thought process in a rubric that you're intimately familiar with.

And then by going through this process manually, you can get an understanding, of this process and help us figure out how we can actually automate and streamline this process on a more systemic level.

And then for individuals that are interested in taking the opportunity even further, you can also help us with the ideation of the product through this rubric calibration process.

You'll be using the platform that we've been developing directly and that'll give you some good insights into potential areas of improvement when it comes to the technology.

And we're absolutely love to hear all of those thoughts.

And for those who want to actually get technically involved and are either comfortable with coding or wanting to learn, they can actually start exploring some of their ideas for automated solutions or passive extraction of this context by actually coding them up and trying to actually build an eval set to test them directly.

But that's the overarching opportunity that we're looking at where you'll be exploring the process of trying to define your own thought process to get a rubric, to be perfectly encapsulate your contextual understanding of a given rubric.

And then while you're doing that, actually trying to figure out a more systemic solution to this AI alignment problem and this problem of trying to define the human thought process when it comes to more ambiguous analysis.

And then when it comes to the actual roles and responsibilities, we have two different opportunities for individuals to explore this project.

In both opportunities, we'll offer some compensation of $500 per fully successfully trained rubric.

However, I do have to note that this isn't amazing compensation.

You'll be able to earn a lot more money per hour as a tutor, because a rubric can take up to 20 hours to train.

So really this is only a project worthwhile if you find or if you see potential value in the actual experience, experience, itself and the potential learnings you can extract out from the process.

And then we have two different opportunities, either more of a casual research project where if you want a little bit more flexibility, and more sort of remote work, then we can explore a more casual research project.

but if you want a very, I guess, immersive experience, you can actually come, and live in the company house, for a minimum of five or so days and sort of work very solidly on this project.

however that is quite an intense, lifestyle.

We have quite an intense company culture, and so that might not suit everyone's schedules and also everyone's actual interests.

when it comes to learning the area, it's largely more tailored for individuals who are very interested in the startup world and want to get exposure and maybe a little bit more mentorship from a yc company, but that's a bit of an overview around us and around the project, if you're interested.

We'd love to hear from you and would love to get your thoughts and support, as we try and do our best to positively impact the education system.