ML engineer focused on generative AI, identity-consistent diffusion, deepfake detection, and graph-based self-supervised learning.
VIPER — Deepfake detection via identity displacement reaction (AB + C → AC + B)
CLIP ViT-L/14 frozen backbone + identity-anchored face preprocessing
0.991 AUC · 95.2% accuracy · 4s per video · trained on 530 videos in 25 min
GitHub · HuggingFace · Live Demo
SynID — Zero-shot identity-consistent image generation from text alone
No real reference photos. No dataset. No pretraining. ~5 min on T4 GPU.
0.954 CLIP identity · 0.791 ArcFace · outperforms IP-Adapter FaceID
GitHub · HuggingFace · Paper
GHOST — Zero-shot supply chain disruption forecasting without labeled data
Closed-loop self-distillation: BiLSTM + dual-level GNN + probe-based drift correction
MSE 0.0081 · 30k+ high-risk cases detected · 180k+ real orders · 15 min on T4 GPU
GitHub · HuggingFace · Paper
PyTorch · CLIP · ArcFace · Diffusers · Hugging Face · ControlNet · GNNs · LSTM · Stable Diffusion
Deepfake Detection · Adapter Training · Bootstrap Learning · Self-Distillation · Drift Detection · Zero-Shot Learning
HuggingFace · ResearchGate · LeetCode · Email