Interested in trustworthy machine learning, physiological AI, language technology, and research-oriented software systems.
Python · Machine Learning · TypeScript · Next.js · Computer Vision · Research Engineering
I’m Susan, an incoming undergraduate student from India preparing to study Artificial Intelligence / Computer Science.
I like working on problems where building the model is only part of the challenge. I’m especially interested in how models fail, how uncertainty should be communicated, how evaluation can accidentally become misleading, and how research ideas can be turned into usable systems.
Most of my current work falls into two directions:
Research-oriented ML
- uncertainty and conformal prediction
- subject variability in physiological signals
- leakage-safe evaluation and calibration
- scientific / biomedical applications of machine learning
Applied AI systems
- language-learning technology
- education and community platforms
- computer vision interaction
- privacy-aware full-stack products
- Can uncertainty guarantees remain reliable for individual subjects when population averages hide difficult cases?
- How does personalization change calibration and selective prediction in physiological ML?
- When does an apparently better model metric hide a worse failure mode?
- How can AI systems expose limitations and provenance instead of presenting every prediction as equally trustworthy?
These questions currently shape much of what I build and read.
|
Subject-conditional conformal prediction for biosignal ML Explores distribution-free uncertainty for ECG, PPG, HRV and EDA models, with particular attention to per-subject coverage rather than population-average coverage. Highlights:
|
Personalized calibration for wearable stress detection Studies whether physiological features should be interpreted relative to each subject’s own resting baseline instead of only through population statistics. Highlights:
|
|
Chemotype-aware uncertainty for molecular docking Combines structure-based virtual screening with scaffold-conditional conformal calibration to study whether pooled uncertainty estimates can fail on specific molecular families. Highlights:
|
Falsification-driven reasoning for LLM research Explores whether uncertain observations can be turned into competing hypotheses, explicit predictions and tests that can genuinely fail, instead of simply producing more persuasive answers. Highlights:
|
|
Community and preparation platform for GKS applicants A full-stack system built around structured applicant discovery, scholarship data, privacy-aware connections and moderation. Technical areas I worked with:
|
Multilingual reading assistant for native web content A Chromium extension for word-level reading support across Korean, Japanese, Chinese and other languages. Technical areas:
|
I’m still early in my academic journey, so I care more about developing good research habits than trying to present myself as an expert.
I try to:
- separate training, calibration and evaluation correctly
- use subject-disjoint or LOSO evaluation when the research question requires it
- inspect failure cases, not only average metrics
- report when an approach makes results worse
- document assumptions and limitations
- keep upstream work and my own contribution clearly attributed
- make experiments reproducible enough that someone else can inspect the reasoning
My goal before and during undergraduate study is to strengthen the fundamentals behind the systems I already enjoy building — machine learning, probability, algorithms, data structures, optimization, statistics and research methodology.
I’m particularly interested in learning from researchers working on trustworthy AI, biomedical / physiological ML, intelligent systems, language technology, and applied machine learning.
Curious enough to build · careful enough to question the result