You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$15,000 left to reach the funding goal.
Building a learning system that remembers how you learn, makes predictions about what you can do, and tests whether those predictions are right.
Working prototype: https://www.youtube.com/watch?v=1JI98fcirps
I started xFilosofem because I kept noticing something strange while studying.
I could know something in one situation and fail to use it in another.
While learning Russian, I could recognize a word immediately when reading it, but struggle to remember it when trying to write a sentence. I could understand a grammar rule, explain how it worked, and still make a mistake when using it in a different context.
I noticed something similar in Medicine.
Knowing the definition of a disease did not necessarily mean I could recognize it in a clinical case. Remembering a laboratory finding was different from understanding why it happened.
These were different kinds of knowledge, but they could easily look the same in a traditional learning system.
A correct answer might increase a score. A wrong answer might decrease it. Eventually, the application would decide that I had learned something and move on.
But had I really learned it?
I started wondering whether a learning system could pay attention to these differences.
What if it remembered not just whether I answered correctly, but how I answered, what mistakes I made, whether I needed help, and what happened when I encountered the same concept in a different situation?
Could it eventually use that history to predict what I would struggle with next?
That question became xFilosofem.
xFilosofem is an experimental learning system built around the idea that learning is a continuous process, not just a collection of completed lessons.
The system keeps a structured history of the learner's interactions.
It records their answers, mistakes, use of hints, and the context in which those things happened. It also keeps track of how answers were evaluated and how that evaluation may change.
Over time, this history can provide evidence about what the learner knows, what they struggle with, and where the system is still uncertain.
Here is a simple example.
Imagine you are learning Russian.
You recognize a word five times while reading. A conventional learning application might treat that word as learned.
But then you try to use it in a sentence and cannot remember it.
xFilosofem should not treat these two situations as identical.
Recognizing a word and producing it are different abilities. The system should preserve that distinction rather than reduce everything to a single score.
Eventually, it should be able to use these observations to make better predictions and choose more appropriate learning activities.
But there is an important condition.
The system must first demonstrate that its predictions are useful.
I do not want to assume that a learning model understands someone just because it can produce a convincing explanation.
I want its predictions to be tested against what actually happens.
Another important part of xFilosofem is its architecture.
I do not want to build another educational chatbot that depends entirely on one AI model.
AI can be extremely useful. It can help evaluate answers, generate exercises, explain difficult concepts, and support richer interactions.
I want xFilosofem to support these capabilities.
But I also want the system's core learning history and continuity to remain independent of them.
An AI model should be replaceable.
A provider should be replaceable.
The learner's history should remain intact.
This means xFilosofem is being designed around its own evidence infrastructure, with AI integrated when it provides a real advantage.
The goal is not to avoid AI. The goal is to use it without becoming dependent on it.
xFilosofem is not just an idea.
We already have a working application.
The project has gone through several stages of architecture, implementation, testing, and revision.
Today, it includes:
A functioning Russian learning domain.
Persistent learner history across interactions.
Records of where learning evidence and evaluations came from.
Revision handling, so previous evidence is not silently rewritten.
Separation between established learning material and generated content.
Architectural support for optional, replaceable AI components.
810 passing regression tests.
A large part of this work is not immediately visible in the interface.
That was intentional.
Before trying to make a system that understands learning, I wanted to build one that could preserve reliable evidence about learning.
If an evaluation changes, I want to know what changed.
If a model makes a mistake, I want to be able to investigate it.
If we replace an AI component, I do not want the learner's previous history to disappear.
The 810 passing tests do not prove that xFilosofem improves learning. They show that we have built and tested a substantial software foundation.
The next stage is where the research begins to test the larger idea.
The first research milestone is deliberately narrow.
Before a learner answers a question, xFilosofem will make a prediction about the outcome.
Then the learner will answer.
We will compare the prediction with what actually happened.
For example, the system might estimate that a learner has a low probability of answering a particular Russian grammar question correctly.
We can then test that prediction.
However, there is another important part.
A complicated model is not automatically a better model.
We will compare xFilosofem's predictions against much simpler approaches.
Perhaps a learner's recent performance is already enough to predict the next answer. Perhaps repetition history or time since the last attempt works just as well.
If a simple approach performs better, I would rather use it.
We also need to examine uncertainty.
A system with very little information about a new learner should not pretend to know exactly what they understand.
Sometimes, the correct decision is to admit that there is not enough evidence.
The first experiment will therefore focus on making predictions before the results are known, comparing them with simple baselines, and examining how reliable those predictions actually are.
Adaptive teaching will come later.
First, we need to know whether the predictions are worth trusting.
Russian is our first domain because it is already the strongest working part of the application.
It also offers interesting learning problems.
A person may recognize a word but fail to recall it. They may understand a grammatical structure but fail to produce it. They may answer correctly in familiar examples and struggle when the context changes.
These differences provide useful opportunities to examine learning evidence.
But I do not want xFilosofem to become only a Russian-learning application.
Medicine is our planned second domain.
And I chose it because it is very different.
In Medicine, learning involves factual recall, mechanisms, interpretation of laboratory results, clinical reasoning, and decisions under uncertainty.
Someone may remember the symptoms of a disease but fail to recognize it in a clinical case.
Someone else may reach the correct diagnosis for the wrong reasons.
These are different problems from learning Russian grammar.
That is exactly why I want to work with both domains.
I want to investigate which parts of learning are specific to a subject and which patterns, if any, can be useful across different subjects.
I am not assuming Russian and Medicine will use the same learning model.
They may need very different models.
But perhaps some evidence about repetition, uncertainty, mistakes, or performance over time can support decisions in both.
Perhaps it cannot.
We need to test that.
This is why using two domains is an important architectural decision rather than simply an attempt to offer more subjects.
There are several ways this work could fail.
The first prediction model may not outperform a simple baseline.
If that happens, I will simplify it.
A more complicated model is not useful simply because it is more interesting.
We may also discover that the system needs much more learner history before it can make reliable predictions.
In that case, it should be honest about its uncertainty.
The Russian model might work well, but the same approach might fail in Medicine.
If that happens, we will investigate what needs to remain domain-specific.
There is also a risk that apparent learning patterns are influenced by which tasks the system selected, rather than by the learner's actual abilities. That needs to be considered during evaluation.
And of course, there are practical risks.
We are a small independent team. I am still studying Medicine, and development time is limited.
We also need suitable evaluation data and independent review to make stronger research claims.
I cannot promise that our hypotheses will be correct.
But I do not think a negative result is necessarily wasted work.
If a model fails, we learn something about its limitations.
If an assumption is wrong, we change it.
I want xFilosofem to be a system that learns from its own mistakes, not one that tries to hide them.
I am building xFilosofem while studying Medicine.
Until now, the project has been self-funded.
I have been able to make progress, but most development happens around my university schedule.
The biggest thing funding would give me is time.
More consistent time to develop the prediction system, run experiments, investigate failures, and document results properly.
Funding would also support hosting, development tools, AI/API usage, evaluation infrastructure, and independent review.
I am a medical student, but I do not want the first Medicine material or evaluation rules to be considered reliable simply because I wrote them.
Outside review will be important.
Rather than making the project dependent on reaching one large funding target, I want different funding levels to support different stages.
Our first priority would be to implement the Russian prediction loop and conduct a small initial evaluation.
We would record predictions before answers are submitted, compare them against simple baselines, and document the results and limitations.
This would not prove the entire xFilosofem concept.
It would give us our first prospective evidence.
This would support more consistent research and development time, a more substantial Russian evaluation, and additional work on uncertainty and prediction quality.
It would also support methodological review and preparation for the initial Medicine domain.
At this level, we could extend the Russian research while developing and reviewing the first Medicine implementation.
The goal would be to begin prospective testing in Medicine and investigate whether any useful learner-evidence structures can be shared between the two domains.
The scale of these experiments will depend on available participants and data.
I would rather complete a small, meaningful experiment than promise a large result we cannot properly evaluate.
I am the founder of xFilosofem and a medical student in the clinical stage of my education.
I lead the learning-system design, research direction, product architecture, and decisions about how the system should represent learner evidence.
My project partner has a background in Electrical and Electronics Engineering. We work together on technical development, testing, and reviewing the application.
He has also worked on technical projects involving computer vision and indoor person tracking, speaker diarization using classical and neural approaches, and multimodal financial decision-support software built with technologies including React and FastAPI.
Neither of us comes from an established research laboratory or a large technology company.
This is my first project of this scale.
I do not have a long history of leading academic research, and I do not want to pretend otherwise.
Our strongest track record is the work we have already done on xFilosofem.
As the project moves into more serious empirical research, I want to involve appropriate external reviewers rather than relying only on our own expertise.
The first outcome I want is straightforward.
A working prediction experiment, a comparison against simple baselines, and a clear account of what happened.
Did the system predict well?
Was it better than simpler methods?
Where was it unreliable?
What should change?
I plan to share the evaluation methodology, baseline definitions, aggregate results, and important limitations where feasible, including negative findings.
xFilosofem itself is intended to remain proprietary software. Sharing research findings should not require exposing private learner data or releasing the application's source code.
The long-term goal is much larger than this first experiment.
I want xFilosofem to become a learning environment that develops alongside the learner, supports multiple subjects and ways of interacting, and makes teaching decisions based on evidence rather than assumptions.
But I do not want to build that future entirely on untested ideas.
When I started xFilosofem, I was mostly trying to understand something about my own learning.
Why could I remember something one day and struggle with it the next?
Why did understanding an explanation not always mean I could apply it?
Why did learning sometimes feel like progress, even when the same mistakes kept returning?
Those questions eventually became a software project.
Then an architectural problem.
And now, a research question.
I have already spent a great deal of time building the foundation for it.
I want to find out whether that foundation can actually tell us something useful about learning.
Maybe our first model will work.
Maybe a simple baseline will beat it.
Either outcome would tell us what to do next.
I am not asking for funding because we have already solved how people learn.
I am asking for funding to help us test whether the system we have built can begin to understand something useful about it.
External funding raised so far: $0. The project has been self-funded.
Working prototype: https://www.youtube.com/watch?v=1JI98fcirps
English is not my first language. I used AI to help translate and edit parts of this proposal. The project idea, architecture, research design, implementation decisions, and claims are mine, and I take responsibility for the final text.