I am a Smart Embedded Systems & IoT undergraduate student at Hanoi University of Science and Technology (HUST), conducting research at the EDABK Laboratory. My work lies at the intersection of Neuromorphic Computing, Deep Learning on Biomedical Signals, and Efficient Edge AI.

• 🎓 B.S. in Smart Embedded Systems and IoT (Expected 2027) @ Hanoi University of Science and Technology (HUST)
• 🔬 Lab Member @ EDABK Laboratory, HUST
• 👨‍🏫 Teaching Assistant @ Global Consumer Intelligence Course, Matsuo-Iwasawa Laboratory, The University of Tokyo
• 💼 Artificial Intelligence Intern @ Viettel Telecom & HANET Technology
• 🏆 2nd Place Winner & Top 100 Global Teams — HSIL Hackathon 2026 (Harvard Health Systems Innovation Lab)
• 🎖️ Outstanding Student — Global Consumer Intelligence Course 2025 (Matsuo-Iwasawa Lab, UTokyo)

• Emerging AI Architectures for Biosignals: Training and deploying Spiking Neural Networks (SNN) and Kolmogorov-Arnold Networks (KAN) alongside neural architecture search (MLP NAS) on biological signals, specifically ECG and PPG.
• Efficient Edge AI: Developing, quantizing, and deploying real-time Computer Vision systems to balance extreme efficiency with high accuracy.

Deep Learning | Spiking Neural Networks | Bio-signal Modeling | Model Quantization

• Concept: Reconstructing clinical-grade chest ECG signals from noisy ear-worn wearable sensors to enable continuous cardiovascular monitoring.
• Methodology: Developed a quantized SNN autoencoder using 4-bit Learnable Step Size Quantization (LSQ) Conv1D layers and MultiSpike activations. Built a composite loss function incorporating MSE, ECGFounder-based perceptual loss, and Pearson correlation.
• Results: Reconstructed signals reached a 0.8508 Pearson correlation on unseen test subjects, rising to 0.9065 with personalized decoder tuning. The network operates on only 13,684 parameters and 13.44M MACs per 2-second window, making it highly viable for neuromorphic edge hardware.
• [Code/Repository]

Computer Vision | NVIDIA DeepStream SDK| TensorRT | Jetson Nano

• Concept: High-throughput real-time product tracking at the edge using resource-constrained devices.
• Methodology: Built a parallelized DeepStream-based video pipeline processing 16 concurrent RTSP streams on a single Jetson Nano. Employed post-training FP16 quantization via TensorRT and knowledge distillation on YOLOv8n.
• Results: Sustained robust detection accuracy and real-time processing constraints at target distances up to 10 meters.
• [Code/Repository]

Real-time Object Tracking | YOLOv8 | Anchor Mapping | Industry 4.0

• Concept: Automating quality assurance in electronics manufacturing lines.
• Methodology: Deployed a custom-trained YOLOv8n detector tracking 11 component classes across a 2-tier packaging setup. Implemented a proprietary "Anchor-based Mapping" algorithm to synchronize state machines across 4 camera feeds.
• Results: Achieved low-latency error alerts for missing components, incorrect packaging order, or positioning errors.
• [Code/Repository]