YufeiMa

From-scratch PyTorch transformer that produces coherent multi-track symbolic music without melody prompts.

• Trained an unconditional symbolic music transformer on the POP909 dataset, applying segmented tokenization to capture phrase-level musical structure across multiple instrument tracks.
• Built a CUDA-accelerated inference loop with configurable temperature and top-k sampling, surfacing the trade-off between conservative continuations and more exploratory outputs.
• Fine-tuned a pretrained checkpoint with a custom tokenizer and 200-token context windows, balancing training throughput against musical coherence over longer horizons.
• Adapted a conditional harmonization framework to the unconditional setting and benchmarked the resulting outputs against the reference distribution.

Dual-sensor IMU pipeline that grades kendo technique on form, force, and exertion in real time.

• Led the backend: a Flask service on Raspberry Pi that ingests time-series IMU and heart-rate data from a Galaxy Watch and a paired shinai sensor, persists sessions to SQLite, and serves a dashboard for review.
• Designed strike-detection and form-scoring algorithms in NumPy — peak detection on accelerometer magnitude with 200ms refractory windowing, plus straightness and consistency scores derived from trajectory variance and waveform correlation.
• Modeled strike kinematics from angular velocity (tip speed v = ω·r) to derive kinetic energy and impact force at the sword tip, then translated raw metrics into actionable training feedback (Effort, Control, Heart).
• Owned the cross-disciplinary integration across REST APIs, embedded sensor sampling, CAD-printed mount hardware, and a small admin endpoint for backfilling new metrics across historical sessions.