Learning by building. Researching for better care.
I explore medical AI, representation learning, and multimodal systems — from respiratory sounds and robust ECG models to organ-centric image segmentation.
Third author. Reproduced temporal and self-supervised baselines, built downstream classification experiments, ran ablations, tuned augmentation settings, and contributed to verification and manuscript revision.
Conditional VQ-VAE → autoregressive prior → controlled synthesis → downstream validation. Dataset 1 macro-F1 improved from 73.20% to 77.15%.
Third author. Contributed to controlled ablation design, robustness evaluation, result analysis, hyperparameter comparison, and manuscript preparation.
Structured masking improved macro AUPRC from 0.7494 to 0.7737 for the compact 1D-ResNet. Gains also held for xResNet1d101 and InceptionTime1d, with consistent direction across all five seeds.
Structured masking · PTB-XL2026
MEDICAL IMAGING03
Ongoing research
OrganSlot / OWT
Modular organ representations
Appendable, optionally frozen organ-specific token banks for modular medical image segmentation.
+10.76 ppDirect-post Dice · WORD ROI20 64.44% → 75.20%
My contribution & method
Implemented PyTorch/DDP training and evaluation, resolved distributed all-reduce deadlocks, and ran 2D/3D shared-head, independent-query, Slot Attention, and reconstruction ablations.
Current evidence is all-class offline. AutoPET Direct Dice was 87.98% versus 88.30% for OWT; indirect reconstruction remains a bottleneck (7.39% on WORD ROI20).
Slot Attention · ViT · DDP2026 — Present
02 / INDUSTRY EXPERIENCE
From experiments to working systems.
Hands-on work across industrial vision, speech recognition, and multimodal model development.
Built PCBA data and evaluation pipelines, explored model calibration and parameter fusion, and implemented U-Net3+ DataMatrix reconstruction with 98% offline accuracy.
Patent contribution
Contributed technical work supporting granted Chinese invention patent CN122510276B on multi-model parameter fusion for PCBA image detection (Sep 2026).
Industrial visionCalibration / ECEU-Net3+
DeepBlue Technology
Nov 2024 — Jun 2025
AI Algorithm and Product Intern
Reproduced CLIP/BERT-based models for image classification, medical NLP, and NER; trained Transformer models and validated Docker deployments.
CLIP / BERTPyTorchDocker
iFLYTEK
Dec 2024 — Feb 2025
Speech AI Algorithm Intern
Prepared and checked speech datasets, assisted Conformer, Paraformer, and Transformer-CTC experiments, and analyzed recognition quality with WER/CER.
ASRAudio augmentationConformer
03 / EXPLORATIONS
Always building something.
Research tools & side projects
MEDICAL IMAGING↗
Lumbar spine imaging
Reproducing the RSNA 2024 lumbar-spine pipeline; training nnU-Net v2 on SPIDER and investigating segmentation-guided anatomical priors.