A year of building a free AI tutor for economics students: Catching up with Tamo Halder

Last summer, friend of the Stone Centre Tamo Halder, Assistant Professor of Economics at Azim Premji University in Bengaluru, India, shared the inspiration behind an AI tutor he was building.

‍

His AI tutor is designed to help first-generation students from disadvantaged backgrounds studying economics, using CORE’s The Economy as a base text. At that point, the tool was at the ideas stage. Almost a year later, Tamo stopped by the Stone Centre again to tell us how far his tutor has come.

‍

Hi Tamo, welcome back. When we last spoke, the tool was just at the ideation stage. Where has it got to since then?
‍

Operationally, we are about a month or two away from getting the tool into a trial phase, where students can actually start testing it. A select group will do an initial round of testing so we can gather feedback, and after that it will move into a more operational, less feedback-driven phase. The larger design and architecture is essentially at the completion stage.
‍

What work is left between here and that trial phase?
‍

The remaining work is largely on the implementation side: how do we host it? Do we buy a separate server for this to run on? Do we host it on the CORE website, or as a separate, independent platform with its own URL? These operational IT questions are what will dominate the next two or three months.
‍

The cost of the underlying AI model itself and the quality of cheaper/free models was a major concern when we last spoke.
‍

It remains the main outstanding challenge, but this connects back to the fundamental principles of the ideation stage. There is already a growing divide in the gen-AI and higher education space between students who can afford to use better models, especially via the paid versions of OpenAI, Anthropic, and the like, versus students who can’t. The quality of response is very different between more expensive models and cheaper ones.
‍

Our goal is to measure quality irrespective of where a student comes from and what kind of paid subscription they have. Our tool is going to be completely free, and the cost of operationalising it sits on the server side, the IT side, not on the user side. We won’t be making API calls to Anthropic or OpenAI for student interactions, which means we have to lean on open-source models, and platforms that run these models.
‍

How are you approaching that technically?
‍

Right now I’m playing around with multiple open-source and free options and trying to fine-tune them. The idea is that even if a particular open-source model isn’t that strong to begin with, fine-tuning it on training data that is specific to our context can make the model tailored to what we need.
‍

On one hand I have an array of open-source models, each with different efficiencies and strengths. On the other, I have a limited window, roughly a two-month period, to pick one and train it well enough to be reliable for trials. So compared with last year, when we were at the ideation stage, it is now a matter of choosing the right model and going ahead with it.
‍

With many models to choose from and limited time, how do you assess whether a model is going to be good enough before committing the time to train and specialise it?
‍

One thing I’ve learned during this process is that testing each response individually by checking the quality of answers one by one takes an enormous amount of time and doesn’t scale. So I am using something simpler as a litmus test.
‍

The design of the tool is such that every student query gets classified first into conceptual, factual, or language-comprehension categories. I’ve noticed that if a model is good at classifying queries correctly, irrespective of the answers it generates in response to student queries, that’s a strong signal that the model will perform well overall.

No model does this perfectly, because some queries are genuinely tricky. But if a model classifies queries correctly about 75% of the time, I’m more confident that the answers it produces will also be reasonable and grounded in the context of CORE. That somewhat intuitive yet ad-hoc correlation is what I’m using to choose between models.
‍

Coming back to the student-side experience, at the CORE Econ workshop at Loughborough in March, you talked about finding the balance between guiding students and not just giving them the answer. Could you say more about where you’ve drawn the limits?
‍

By the very principles of how I am designing the tutor, which involves complementing a student and not substituting their entire learning process, I can’t allow for an endless back-and-forth between the AI model and the student. On top of that, no model, including paid tiers of GPT and Claude, can maintain context indefinitely; the context gets weaker as the chat goes longer. So both ways, I’ve had to deliberately curtail the conversation that happens between the student and the tutor.
‍

That gives you a limited window in which the AI model has to do its job, which is twofold: see where the student is struggling, whether with a concept, a fact, or language, and then provide the right kind of clarification within that window.
‍

Learn more about the 2026 CORE Workshop, including the discussions about AI and economics, in this recap on the CORE website.
‍

How’s the instructor side of the tool shaping up?
‍

The instructor side of the dashboard has come a long way. Instructors now get information on who is struggling, where, and what kind of queries a student has: language-based, conceptual, or factual. With that, an instructor can do two things.
‍

First, if a particular student is repeatedly sending queries of a certain kind, the instructor can offer targeted support. Language support, for example, or pointing them to other resources outside of economics.

Second, if they see that, say, 70% of the class is getting stuck on a particular concept, they can adapt their next lecture before moving on to new content. On top of that, some advanced analytics will be available for instructors that give them some aggregate information on how much attention students are paying while reading the textbook.
‍

That’s why we see the role of the AI tutor as something that doesn’t take the human element away from either the student or the instructor. The design is often restrictive, and you have to work within that. We are positioning the tutor as a complement, not a substitute.
‍

Thank you for the update, Tamo. We look forward to hearing how the student trials go.
‍

Read our previous interview with Tamo here.
‍

CORE’s The Economy, used for the development of Tamo’s AI tutor, is a free, open-access economics textbook.