Experience

Implemented the Vision Transformer (ViT-B/16) architecture from scratch in PyTorch, following the paper "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." Manually built all core components, including convolutional patch embeddings, class and positional embeddings, Multi-Head Self-Attention (MSA) and MLP blocks with Layer Normalization (LN) and residual connections, as well as the final classification head.

Used the equations and architectural definitions from the original paper to reason about data flow and tensor transformations throughout the model, explicitly tracking tensor shapes step-by-step from input images to output classification in order to ensure correctness and deepen understanding of the model structure.

Validated the implementation end-to-end by training the model from scratch on a 5-class weather image classification dataset sourced from Kaggle. Documented training simplifications relative to the paper and compared the custom implementation with PyTorch’s built-in ViT.

As a senior member of team Schwarzschildren, conducted in-depth theoretical and experimental research on open-ended physics problems, winning a gold medal in 2023 and a bronze medal in 2024. Researched various physics topics including waves, acoustics, electromagnetism, collision theory, beam theory, fluid dynamics, and optics.

Developed simulations, designed experimental setups, and collected experimental measurements; conducted data cleaning and statistical analysis, interpreting results to validate physical models using MATLAB and Python.

Constructed research reports, presented, and defended findings in formal physics debates, demonstrating strong scientific reasoning and communication.

Conducted controlled measurements to research key pendulum relationships: Period vs Angle, Period vs Length, and Q-factor vs Length. Designed and fabricated a custom experimental apparatus to systematically investigate pendulum characteristics.

Developed a Python-based analysis pipeline in Jupyter Notebooks to extract motion data from Tracker and quantitatively compare results with theoretical predictions. Identified and characterized deviations from standard models at large angles and with non-negligible rotational inertia, proposing a refined pendulum model to account for nonlinear and damping effects.

Conducted research to identify optimal matboard box beam bridge designs using a custom Python structural analysis and simulation tool developed in Google Colab.

Simulated loads and computed Factors of Safety (FOS) across design candidates to evaluate and compare structural performance. Validated theoretical predictions by constructing a physical matboard bridge, achieving a maximum load of 640N, exceeding the 400N baseline by 60%.

Produced a 4000-word research paper integrating empirical data and literature to assess autonomous vehicle (AV) societal and regulatory impacts. Researched and analyzed deep learning methods for AV perception and path-planning, evaluating their technical and ethical limitations. Conducted a primary survey on public AV perceptions, generating quantitative insights on safety and ethics.

Researched and developed a mathematical thrust-curve model for black-powder rocket engines that reflected the propagation of 3D flame-front geometry in a two phase burn. Developed a Python program that extracted rocket motor datasheet features, optimized model constants, and batch-processed CSV outputs.

Achieved low median prediction errors for average thrust (13.42%), maximum thrust (15.07%), and tailing thrust (9.31%), while identifying limitations in total impulse and burn-time predictions.

Conducted experimental research on cantilever dynamics, testing the hypothesis that increasing length decreases stiffness and lowers natural frequency (f ∝ 1/L²).

Analyzed sound waveforms in Audacity by applying noise gating, spectrogram analysis, and Fourier transforms, then performed curve-fitting in LoggerPro to validate the predicted inverse-square relationship.

Investigated the relationship between alcohol chain length (methanol to octanol) and soot production, testing the hypothesis that alcohols with higher molar enthalpies of incomplete combustion produce more soot.

Designed a dual-purpose calorimetry and soot-collection experimental setup to measure heat release and capture soot under controlled airflow. Processed data with Vernier Graphical Analysis and Microsoft Excel, observing a strong linear correlation between soot mass per mole and molar enthalpy (R² = 0.986).