Hi everyone, would appreciate your comments and advice on below motivation letter for Erasmus Mundus Scholarship
-------------------
As a founding team member at an AI-powered mental wellness startup, I have confronted a technical challenge I did not fully anticipate when I began: aligning real-world clinical expertise with the model we had built. We had a base model and a working dataset pipeline, and we had mental health professionals on the team defining techniques such as Neuro-Linguistic Programming and Cognitive Behavioral Therapy, but translating that expertise into something the model could reliably learn and apply proved far harder than integrating any standard feature or dataset. The techniques were qualitative, context-dependent, and grounded in professional judgment, while our model needed structured, learnable signal, and closing that gap exposed a limit in my own knowledge.
I could engineer around problems I understood, but I did not yet have the theoretical and research grounding to reason about a mismatch this fundamental between expert knowledge and model architecture. That realization is what led me to apply for the EMAI programme: not to keep working with models as they are handed to me, but to understand the structure and foundations behind them well enough to shape what they do, rather than simply use what they already offer. I want this programme to move me from being a user of models to someone who understands the theory, architecture, and infrastructure behind state-of-the-art AI, and the open problems researchers around the world are working to solve.
My academic interest in this area began with my Bachelor's final-year project, where my group built an AI system to support speech and communication development in children with Autism Spectrum Disorder. A significant part of the work was going through existing literature on speech modeling for neurodivergent populations. It was here that I first experienced research as a discipline distinct from engineering: the system had to be developed in close consultation with professionals working in the domain, whose clinical understanding of ASD communication patterns shaped decisions we would otherwise have made purely on the data. That process taught me that a working model and a correct one are not the same claim. Models like word2vec, built to align input with output through patterns learned from typical speech, strained against the far less predictable rhythms of neurodivergent communication, where the same underlying intent could surface through very different, non-standard cues. Designing feedback generation compounded this: the system needed to critique a child's communication attempt without it reading as failure, which meant the model's output had to carry not just accuracy but tone, and tone is not something a loss function optimizes for on its own. Working through this, closely and iteratively with domain professionals, taught me that building a system that works and building one that understands the human tendencies behind its own outputs are, in research, two separate problems
My interest in Artificial Intelligence started earlier than either of these experiences. In coursework, I first implemented preprocessing pipelines and standard learning algorithms on real datasets and watched a model actually learn. There is a particular moment in that process that hooked me: setting up a loss function, watching it descend across epochs, and realizing that a handful of mathematical operations, repeated and corrected against error, was enough to produce a decision boundary I had not hand-coded. Validating that boundary against a held-out test set, and seeing where it generalized and where it quietly overfit to noise in the training data. That instinct is what carried me from isolated notebooks to building systems: taking a model out of a controlled experiment and into a pipeline that ingests real, messy data, serves predictions, and has to keep improving as more data arrives. Each step of that progression, from a single trained classifier to an end-to-end platform, sharpened the same underlying fascination, not with AI as a tool that already works, but with the mathematics and structure that let it learn in the first place. It is this fascination that convinces me to pursue studies in this field.
The chance to study within and across four distinct research cultures, rather than one, matches how I want to grow: exposed to different methodological traditions, different open problems, and different ways of framing what "good AI" means, rather than absorbing a single school of thought. Within the programme, I am especially drawn to the Data Science track, given my background in extracting signal from noisy, human-generated data, though the human-centered questions raised by Intelligent Decision Making are equally close to my interests. I want to work directly alongside researchers actively publishing in human-centered and assistive AI, in a setting where I can test my ideas against people who will actually challenge them - something that was structurally difficult given the research infrastructure available to me at home.
What I believe I can contribute to an EMAI cohort is a perspective grounded in deployment as much as theory: I have shipped software under real constraints, worked inside an AI product facing real users, and pursued an underserved research area that combines technical and social urgency. I am not approaching this programme as an abstract next step, but as the continuation of a specific question I already care about and the infrastructure I currently lack to pursue it properly.
I am seeking, through EMAI, the mentorship, research environment, and international academic exposure to move from someone who has built AI systems to someone equipped to meaningfully contribute new knowledge to the field, particularly at the intersection of AI, data science, and human communication.
-------------------
As a founding team member at an AI-powered mental wellness startup, I have confronted a technical challenge I did not fully anticipate when I began: aligning real-world clinical expertise with the model we had built. We had a base model and a working dataset pipeline, and we had mental health professionals on the team defining techniques such as Neuro-Linguistic Programming and Cognitive Behavioral Therapy, but translating that expertise into something the model could reliably learn and apply proved far harder than integrating any standard feature or dataset. The techniques were qualitative, context-dependent, and grounded in professional judgment, while our model needed structured, learnable signal, and closing that gap exposed a limit in my own knowledge.
I could engineer around problems I understood, but I did not yet have the theoretical and research grounding to reason about a mismatch this fundamental between expert knowledge and model architecture. That realization is what led me to apply for the EMAI programme: not to keep working with models as they are handed to me, but to understand the structure and foundations behind them well enough to shape what they do, rather than simply use what they already offer. I want this programme to move me from being a user of models to someone who understands the theory, architecture, and infrastructure behind state-of-the-art AI, and the open problems researchers around the world are working to solve.
My academic interest in this area began with my Bachelor's final-year project, where my group built an AI system to support speech and communication development in children with Autism Spectrum Disorder. A significant part of the work was going through existing literature on speech modeling for neurodivergent populations. It was here that I first experienced research as a discipline distinct from engineering: the system had to be developed in close consultation with professionals working in the domain, whose clinical understanding of ASD communication patterns shaped decisions we would otherwise have made purely on the data. That process taught me that a working model and a correct one are not the same claim. Models like word2vec, built to align input with output through patterns learned from typical speech, strained against the far less predictable rhythms of neurodivergent communication, where the same underlying intent could surface through very different, non-standard cues. Designing feedback generation compounded this: the system needed to critique a child's communication attempt without it reading as failure, which meant the model's output had to carry not just accuracy but tone, and tone is not something a loss function optimizes for on its own. Working through this, closely and iteratively with domain professionals, taught me that building a system that works and building one that understands the human tendencies behind its own outputs are, in research, two separate problems
My interest in Artificial Intelligence started earlier than either of these experiences. In coursework, I first implemented preprocessing pipelines and standard learning algorithms on real datasets and watched a model actually learn. There is a particular moment in that process that hooked me: setting up a loss function, watching it descend across epochs, and realizing that a handful of mathematical operations, repeated and corrected against error, was enough to produce a decision boundary I had not hand-coded. Validating that boundary against a held-out test set, and seeing where it generalized and where it quietly overfit to noise in the training data. That instinct is what carried me from isolated notebooks to building systems: taking a model out of a controlled experiment and into a pipeline that ingests real, messy data, serves predictions, and has to keep improving as more data arrives. Each step of that progression, from a single trained classifier to an end-to-end platform, sharpened the same underlying fascination, not with AI as a tool that already works, but with the mathematics and structure that let it learn in the first place. It is this fascination that convinces me to pursue studies in this field.
The chance to study within and across four distinct research cultures, rather than one, matches how I want to grow: exposed to different methodological traditions, different open problems, and different ways of framing what "good AI" means, rather than absorbing a single school of thought. Within the programme, I am especially drawn to the Data Science track, given my background in extracting signal from noisy, human-generated data, though the human-centered questions raised by Intelligent Decision Making are equally close to my interests. I want to work directly alongside researchers actively publishing in human-centered and assistive AI, in a setting where I can test my ideas against people who will actually challenge them - something that was structurally difficult given the research infrastructure available to me at home.
What I believe I can contribute to an EMAI cohort is a perspective grounded in deployment as much as theory: I have shipped software under real constraints, worked inside an AI product facing real users, and pursued an underserved research area that combines technical and social urgency. I am not approaching this programme as an abstract next step, but as the continuation of a specific question I already care about and the infrastructure I currently lack to pursue it properly.
I am seeking, through EMAI, the mentorship, research environment, and international academic exposure to move from someone who has built AI systems to someone equipped to meaningfully contribute new knowledge to the field, particularly at the intersection of AI, data science, and human communication.
