GlassBoxML — A machine learning library built from scratch in NumPy, with no scikit-learn under the hood. Implements linear regression, logistic regression, decision trees, KNN, and random forests (with bootstrap aggregation and out-of-bag scoring), all covered by a 31-test pytest suite. Includes an interactive FastAPI + Chart.js webapp for training and visualizing each model live in the browser, fully containerized with Docker and deployed to Kubernetes via minikube, with CI/CD running in parallel on both GitHub Actions and a self-hosted Jenkins pipeline.
Resource-Efficient Neural Networks (Quantization) — A research project investigating fixed-point quantization for neural network compression, covering post-training quantization (PTQ), quantization-aware training (QAT), and a custom differentiable quantization method. Identified and characterized "gradient deadlock" — a failure mode where weight updates smaller than the quantization step get erased by rounding — and tested gradient scaling and explicit error accumulation as fixes. Designed a parameterized differentiable stair non-linearity with a curriculum-based annealing schedule, letting models train through hard quantization constraints and ultimately outperform the standard PTQ baseline. Experiments progressed from linear regression (isolating how input conditioning affects quantization error) through non-linear targets to a two-layer neural network, analyzing how bit allocation and numerical conditioning shape quantized model accuracy.
Echoes of Longevity — Healthy Ageing Narratives in Science, News, and Social Media A research project studying how ideas about healthy ageing, longevity, and biohacking circulate and evolve across scientific, journalistic, and social media discourse from 2010 to 2025. Built a cross-platform corpus spanning PubMed, news outlets, Reddit, and YouTube, and applied topic modeling, transformer-based sentiment and emotion analysis, interrupted time-series methods, and Granger causality to trace how narratives emerge, spread, and shift over time. Found that vocabulary follows a consistent "mechanism-to-action" gradient — from PubMed's clinical language to news's human-interest framing to YouTube's advice-driven tone to Reddit's community-protocol style — while sentiment converges toward cautious optimism across all platforms. Identified selective cross-platform lead-lag effects (e.g., YouTube coverage anticipating later news reporting on cardiovascular disease) and showed that COVID-19 acted as an emotional turning point that normalized rather than diminished public engagement with longevity topics. Won 1st Place with Special Recognition at BME's TDK Student Scientific Conference and 2nd Place at EELISA.
