About
I gravitate toward machine learning problems where signals are noisy and observations are incomplete. Across my research, the same methodological questions continually arise: what can be learned from sparse or difficult observations, how stable that learned structure remains across users and conditions, and how far representation design and temporal modeling can compensate for what the data or hardware cannot provide.
Across these projects, the model could not be separated from the conditions that produced the data: the sensing setup, the representation strategy, the evaluation design, and the eventual deployment context.
Viewing Machine Learning Holistically
This systems perspective led me to design evaluations around generalization, subject-level separation, and the full preprocessing-to-modeling pipeline, rather than treating performance on a held-out split as the only measure of success.
To support this kind of work, I developed BioHCI, a 10,000+ line modular time-series learning framework written in Python that evolved from early physiological modeling, through capacitive touch sensing, into reusable infrastructure for structured human-generated signals. BioHCI allowed me to treat modeling, preprocessing, evaluation, and deployment as reusable and interdependent parts of the same learning system instead of as unrelated stages, custom built for each project.
This systems mindset carries a less obvious benefit: it makes the work more interpretable, not just to reviewers, but to the engineers, clinicians, and designers who need to act on what the model produces. Interpretability allows for modeling decisions that are analytically defensible, legible to domain experts, and implementable on the systems that actually exist.
I typically treat predictive tasks not only as endpoints, but also as tests of whether the representation captures the structure the task depends on: does a model perform well, and also does it capture structure that is stable across users, interpretable in context, and valid beyond the original evaluation setting? Representation quality, generalization under shift, and learning from limited or noisy signal are the same problems that make fine-tuning and deploying foundation models outside controlled conditions difficult, which is part of why this work extends beyond the domains in which it was developed.
Research Background
I earned my PhD in Computer Science at Drexel University, where my research spanned both physiological signal modeling and interactive textile sensing. My early work included fNIRS data, which helped build my foundation in signal processing, feature construction, and machine learning for noisy, temporally structured human data. That work also laid the foundation for BioHCI, the modular framework I later expanded across multiple sensing domains. It also introduced questions that would stay with me: how temporal resolution affects what can be learned, how representation choices shape performance, and how much useful structure can be recovered from difficult biological signals.
My dissertation focused on the real-world interactivity potential of minimalistic knitted sensors at the intersection of machine learning and human-computer interaction. Building on my earlier experience with physiological signals, I developed computational approaches for sparse textile-based sensing systems that could recover human intent through touch and gesture without sacrificing manufacturability, robustness, or usability.
Following my PhD, I completed a postdoctoral fellowship at the Center for Autism Research at CHOP, where I developed computational frameworks for behavioral and physiological time-series data in naturalistic settings. Across these stages, the central problem has been the same: recovering stable, meaningful structure from noisy, human-generated signals and evaluating whether that structure holds beyond the conditions in which it was first observed.
What I Bring
The specific domains of my published work — physiological signals, textile sensing, and behavioral analysis — are narrower than the range of machine learning problems I can contribute to. The reasoning transfers across domains, as do specific technical strengths in representation learning, temporal modeling, cross-subject evaluation, and end-to-end machine learning for structured and sequential data:
- designing representations for signals that are sparse, noisy, temporally structured, and variable across users and conditions
- building temporal models and evaluation frameworks that test for generalization rather than only optimizing held-out accuracy
- making architecture decisions with deployment and implementation constraints in mind from the start
- working across research, engineering, and applied contexts without losing rigor in either direction
Technically, this work has involved Python, PyTorch, scikit-learn, NumPy/SciPy, sequence models including LSTMs and CNNs, and edge deployment on NVIDIA Jetson platforms.
I work well in research settings where the right model, representation, or evaluation strategy is still an open question, and in engineering settings where the challenge is making a trained system work reliably in practice. The most productive problems, in my experience, sit precisely at that boundary.
Collaboration
Across my work, meaningful collaboration has taken different forms: translating user study findings into concrete modeling requirements with HCI researchers, making evaluation design legible to clinicians who needed to trust the output without understanding the pipeline, and working with engineers on the gap between a system that runs on embedded hardware and enough signal to understand human intent.
In each case, the technical and the contextual had to develop together: domain knowledge shaped which modeling choices were defensible, and modeling constraints shaped which research questions were actually answerable.
Explore further
Research: Questions, methods, and findings across three research domains, with the modeling reasoning made explicit.
Systems: BioHCI and the end-to-end pipeline architecture behind the work, from signal acquisition through embedded deployment and robustness evaluation.
Publications: Full publication list organized by research area, with abstracts, PDFs, and slides.