Junior CS @ NJIT | NASA-Funded ML Researcher | GPU Computing & Deep Learning
Building production-grade ML systems from CUDA kernels to transformer architectures.
NASA MIRO Research Fellow | Dec 2025 - Present
Training multimodal transformer models on decades of NASA/SDO and SOHO solar data to predict eruption events. Achieved 20% forecast accuracy improvement through systematic experiment tracking and Physics-Informed Neural Network integration with YOLO-based feature extraction.
Open Source
Contributing to Hugging Face Transformers — PyTorch CUDA backend API migrations (PR #45856)
Implemented Flash Attention (Dao et al. 2022) forward and backward passes from scratch in raw CUDA C++.
Performance:
- 22.9× softmax speedup (4.8 → 110 GB/s) via shared memory and warp-shuffle reductions
- 86× HBM traffic reduction at seq_len=4096
- Verified all gradients (dQ, dK, dV) against PyTorch autograd with max error <4e-7
Techniques: Shared memory tiling, warp-level primitives, online softmax, gradient recomputation
Technical blog post | Benchmark results
Recall — Edge-Deployed Facial Recognition
Winner: Best Use of Grok (XAI) & Best Use of Arm | HackPrinceton 2025
Edge-deployable face recognition pipeline built in PyTorch and OpenCV for Raspberry Pi.
- <200ms on-device inference latency
- 94% precision with <2% false positives
- Streaming REST API with OpenAPI docs
Stack: Python, PyTorch, OpenCV, FastAPI, Raspberry Pi
ROM-COM — Stroke Rehabilitation System
HackPrinceton Spring 2026
Real-time gesture recognition for physical therapy assessment.
- 96%+ CV accuracy using MediaPipe and Random Forest
- Live FMA-UE clinical scoring from joint kinematics
- 14+ FPS WebSocket streaming with sub-50ms latency
Stack: Python, MediaPipe, scikit-learn, FastAPI, React
AI/ML: PyTorch, TensorFlow, Hugging Face, OpenCV, CUDA, cuDNN
Languages: Python, C++, CUDA C, JavaScript/TypeScript, Java, SQL
Tools: Docker, Git, FastAPI, Flask, AWS, Azure, MongoDB
- HackPrinceton 2025 — Best Use of Grok (XAI) & Best Use of Arm (MLH)
- NASA MIRO Fellow — Competitive research fellowship for ML in heliophysics
- Apple Technical Specialist — Top 5% nationally for quality and efficiency
LinkedIn: andreayanezsoto
Website: andreasoto.dev
Email: andreayanez11@outlook.com
Open to ML research collaborations and open source contributions.
