Projects

A fast waveform-level attack that trains a residual 1D U-Net to generate small, bounded audio perturbations in one forward pass, keeping speech understandable while sharply degrading Whisper transcription.

• Result: Whisper Small WER rises from 0.047 to 3.825 with 99% attack success.
• Transfer: the same Small-trained generator remains effective on Whisper Medium and Large-v3 without retraining.

Exam project combining reinforcement learning and LLM-based guidance for navigation in a stochastic PyBullet environment with a hidden goal, sparse local sensing, dense guidance embeddings, and recurrent control.

• Work: designed the LLM oracle interface and constrained structured outputs.
• Outcome: reached 62% held-out success while exposing robust execution as the main bottleneck.

A computer-vision pipeline for large drone images where the goal is simple: spot unusual objects or target areas quickly, then turn them into regions that can be checked or classified.

• Task: find candidate target areas inside high-resolution aerial frames.
• Why: targets are rare and images are large, so narrowing the search saves time.
• Output: heatmaps, bounding boxes, and crops that can feed the next classifier or demo.

A small forecasting study on a chaotic system: can a model predict the next movement, and how quickly do tiny mistakes grow when predictions are fed back into the model?

• Task: predict short pieces of motion in a system that is very sensitive to small changes.
• Why: it is a clean toy problem for understanding the limits of time-series prediction.
• Lesson: good short-term predictions do not automatically stay reliable over long rollouts.

Exploratory analysis of the 2019 European elections: data cleaning, hypothesis checks, and visual summaries.

• Work: wrangling messy tables, checking consistency, and building reproducible notebooks.
• Output: plots and narrative conclusions based on quantitative patterns.

A first-person Three.js experience that turns sound into a place you can explore: a live waveform shows the signal over time while an audio-reactive skyline reveals its frequency spectrum.

• Visualization: a 70-band FFT drives building height, brightness, and ground flow.
• Interaction: real-time sources, filters, pitch controls, and a tracking robot arm connect audible changes to visible feedback.

Small synthetic-data pipeline for generating labeled images with domain randomization for YOLO-style training.

• Work: scene randomization, lighting changes, and label export.
• Why: fast iteration when real data is expensive, limited, or hard to collect.

Lightweight backend/API work and small apps focused on turning functionality into something usable and testable.

• Work: API schemas, endpoints, and clean integration with client logic.
• Stack: OpenAPI, SQL fundamentals, HTML, CSS, and JavaScript.

A growing space for browser-native simulations, visual explainers, and small tools that are useful to explore but not necessarily large enough for standalone project pages.

• Task: keep small interactive ideas discoverable as they evolve.
• Why: simulations and sliders can explain some concepts better than static text.