Rambling along my journey, because sometimes you need to let yourself drift.
The days went by, and I was still mentally thrown off by having landed without the right maps. But I had to adapt however I could and enjoy the journey.
My technical architecture evolved, and by that I mean it pivoted drastically, several times.
Seeing the complexity required by the activities and, in general, by all the data, the possibility of finishing the summer with my own trained model became more and more unreachable as the days went by. Not only because of the amount of activity-based data I had to create, but also because of the subtle nuances I needed, which were in turn related to the needs of students in specific contexts and had to be incorporated into the model if it was truly going to be different from "any other language model". What a few weeks earlier had seemed perfectly feasible suddenly began to look like a difficult task for months, and impossible within the timeframe of the scholarship.
During the following mentoring sessions, we decided to turn toward an agent system in which the necessary guidelines for each agent, their hierarchy, and so on, had to be defined with great care.
This led me to deconstruct my own thinking process so I could understand how, in my previous professional life, I created adapted activities in which everyone could participate.
During this process I became aware that, while the LLM constantly recommended something like this:
1. Recognize the students' needs.
2. List the support needs.
3. Register the tasks.
4. Compile the activity.
My thinking process was completely different:
First, I decide what activity I want to carry out.
Then I break down the tasks.
Then I distribute them while already taking needs and adaptations into account.
If there is an adaptation that needs to be made specifically for a concrete context, I look for materials, supports, or add small changes so that the task is sufficiently adapted.
At the same time, I began working with local Ollama models. This consumed a lot of resources on my device, so we even set up a server at home with an old computer to serve them to me. But while developing with CrewAI and Ollama, I started running into problems. I also tried other agent frameworks, such as LangGraph or BeeAI, but I could not get them to work well.
The times when I managed to tie the threads together and get answers, I realized that the models do not know how to "hierarchize" tasks, and for me this part of the process was essential. The tasks and subtasks before carrying out a game, for example, involved creating the materials for the game themselves, and that very process of creating the materials is already learning and pedagogical work. So I worked hard to find a solution to this.
After testing and testing things, I ended up exploring a hypothesis for incorporating a "quantum processing layer", yes, you can laugh, but the way I saw it was:
"If I consider the classroom as a complex system made up of active/inactive individuals, the students, perhaps the solution could involve creating a new system that emerges in a certain direction, with that distribution, toward a higher state".
I investigated to what extent a "quantum processing layer" could solve this, or whether I was losing the plot again.
I asked Ismael for help to find out whether this made any sense at all. You can probably imagine the answer: I was taking a stroll around Saturn. I have to admit that I was beginning to feel sorry for my mentors, given the patience they were showing me!
I went back to the struggle between CrewAI and Ollama and finally said to myself:
If an agent is made up of an agent, a task, and a tool, it cannot be that hard to make them "by hand".
So I rolled up my sleeves and stepped into the mud up to my knees. I began compiling my agent system without a framework, yes, I admit it, that is what I did. It seemed that I was "getting something", even if it was not very high quality, so the next step was to incorporate "human in the loop" to help the model generate more accurate answers.
I worked on how to enrich the agents' instructions so that the answers would be specific and concrete for the task, but I ran into difficulties again in that regard.
Sometimes the journeys you had planned turn upside down, so you have to stop at a terrace and look at the maps again. But I still had days left, and I had to plan as well as possible under the circumstances I had.