Letters / Mot. Letter - ETH Zurich - Masters in Data Science - regarding intro & narrative [4]

Prompt:
Approximately one A4 page in the language of instruction of the chosen programme, stating your motivation for choosing this specific Master's degree programme at ETH Zurich.

Your letter of motivation is probably the most personal document in your application. It gives you an opportunity to introduce yourself to the degree programme, and to explain your motivation for conducting Master's studies at ETH Zurich. The letter is your chance to provide relevant and interesting insights about yourself for the admissions committee of the respective Master's degree programme, and to show why you are the right person for that degree programme.

In the letter we also expect you to demonstrate how familiar you are with the desired degree programme and its characteristics. The letter should describe your interests, your competences and your motivation for choosing the specific degree programme at ETH Zurich. It should also set out your expectations as to teaching, and your future plans.

Content:

For eight years, my father struggled to secure a loan to start his dream business. Banks considered him not creditworthy, despite his intent to repay. When he finally got a loan, he proved them wrong by paying on time. However, his humiliation already cost our family; I missed educational opportunities because of systems neither of us could understand. Early in my undergraduate studies, I was motivated to solve this through AI. But after joining the credit industry, I realized that using machine learning for loan approvals presents its own complex challenges: it must be performant, robust, and, most importantly, explainable.

At XXX, based on counterfactual loans, I discovered that 24% of rejections were of creditworthy individuals. They simply lacked traditional bureau records, and our statistical models could not capture enough signals from unstructured SMS data. To address this, I built a neural network model with the potential to unlock about $50 million in loan disbursements annually. However, stakeholders raised concerns about its stability and black-box nature, as it's not ethical to reject any user without clear explanations. Even SHAP values only help explain the influence of features on credit risk, but not their connection to users' behavior. At that moment, I realized I had built the same opaque system that led to my father's unclear loan rejections. Watching a better model fail to reach customers was disheartening; I asked myself: "Can I build models that capture complex financial behaviors while providing clear explanations that customers like my father can trust?" This question motivated me to pursue a Master's in Data Science at ETH Zürich, where I aim to research robust and interpretable neural networks to develop trustworthy AI models that enhance financial access for underserved populations.

My journey toward reliable AI began during my undergraduate studies at VIT University. For my capstone project under YYY, I developed an Instruction-aligned LLM Tutor to guide students through problem-solving rather than giving direct answers. I trained small open-source LLMs that were vulnerable to prompt injection attacks. To mitigate this, I created a synthetic adversarial dataset to improve robustness, but the model remained vulnerable to such issues. Moreover, the difficult part was that I could not pinpoint the reasons for its failures. Despite these challenges, this system served over 200 students, earning me the "Best Capstone Project Award of 2024." This project taught me a valuable lesson: AI robustness and explainability are intertwined. Systems that cannot be explained cannot be made robust. The issues of model robustness and explainability became even more critical in my professional work. In an applied research project, I built a cash-flow extraction model that processes SMS messages to understand users' financial behavior across 14,000 vendors. The main challenge was generating diverse training data to develop a robust model. I processed a dataset of 100 million messages. Despite this, about 3% of samples remained misclassified after every model iteration. To understand why misclassification occurred, I used integrated gradient techniques to identify patterns where the model underfitted or overfitted. This allowed me to strategically improve the training data, enabling the model to operate stably in production for over a year.

ETH Zurich offers the rigorous environment to advance this research, and Professor XXX exemplifies the mentorship I seek. His work on reliable and interpretable AI directly addresses the core challenge I faced at XXX: building models that are both accurate and trustworthy. Professor Vechev's research on combining certified robustness with explainability offers what my deployed neural network lacked. His techniques for proving bounds on model behavior would allow me to guarantee that credit scores stay stable under input variations while providing interpretable explanations for each decision. His course "Reliable and Trustworthy Artificial Intelligence" covers adversarial robustness, verification methods, and interpretability techniques. Working with his research group and the ETH AI Center's infrastructure would enable me to develop credit scoring architectures where every component has both certified properties and clear explanations.

For my thesis, I propose developing decomposable credit scoring architectures that convert unstructured data into a risk score along with a human-readable explanation. The interpretable sub-component (e.g., "income stability: 0.7, expense predictability: 0.8") and ability to attribute specific data points to the credit score will enhance model explainability.

After completing my Master's, I aim to pursue applied research at institutions bridging academia and industry, focusing on developing explainability frameworks for financial services. Having worked in Indian and African markets, I have seen how millions of people are excluded due to technology's current limitations. My goal is to ensure that creditworthy individuals like my father are never turned away because of opaque algorithms. ETH Zurich provides the exact technical depth and research environment I need to make this vision a reality.

Questions:
- I would appreciate an overall assessment of my motivation letter.
- The personal motivation is currently described across two paragraphs; should it be shortened further?
- My proposed thesis direction feels slightly outdated given my current understanding. Would it be better to present it more broadly?
- Since I don't have published research, I briefly described my technical work. Should I simplify this section instead of detailing specific projects, especially given the one-page limit? If so, would adding only selected impactful projects be more appropriate?

Note: I used AI to polish the sentences in my motivation letter. I have rewritten mutliple times, but couldn't able to reduce the AI detection. Not sure what to do