徐仕杰 · Building AI with peace, love & curiosity. Don't just read about me — ask my AI below.
I'm a Data Science graduate from XJTLU (Rank 2/241, First-class Honours) heading to Columbia University for an M.S. in Artificial Intelligence. I like turning research ideas into systems that actually run.
My work sits where multimodal machine learning meets applied decision-making — LLMs and vision-language models for medical AI, reinforcement learning for real-world decisions, and trust/consensus modeling for large-scale group decisions. Across these I co-authored 7 papers (including IEEE TFS, ICASSP and ICME).
I'm easy-going but relentless, endlessly curious, and I try to build with a bit of peace & love. Lately I'm hands-on with vibe-coding tools — shipping small, complete AI things and learning in public.
Causal uplift modeling to optimize default content display on Baidu's Discovery feed.
A resource-efficient vision-language framework for trustworthy skin-cancer diagnosis.
Uses an LLM as an explicit causal reasoner over a frozen SAM encoder for robust cross-domain medical segmentation.
End-to-end framework unifying multimodal sensing, yield forecasting and resource decisions with reinforcement learning.
A lightweight diagnostic system for ICU/ER lung ultrasound, validated with a hospital blind test.
Hybrid statistical + deep time-series forecasting of seasonal power load.
Reaching consensus in large social-trust networks via structural-hole spanners and dynamic trust.
Multi-label lumbar-spine degenerative classification — Silver medal (Top 5%).
Robust object detection in heavy fog and low light for safer autonomous driving.
A fused statistical/ensemble framework for dynamic prediction, applied to Olympic-medal forecasting.
A bi-level consensus framework using structural-hole spanners + fuzzy trust updates; lifts group consensus 0.589 → 0.826.
LLM as a causal reasoner over a frozen SAM encoder for cross-domain robustness — strong OOD gains with <8% trainable params.
Frozen CLIP + quantized Qwen2.5-VL (LoRA) with a Consistency-aware Focal Alignment loss; beats 13B baselines by 4.3–6.2% acc, 43% fewer params.
Hypergraph modeling of higher-order group interactions; faster empirical consensus convergence than graph-based models.
End-to-end 3D-CNN + TCN + Dueling DQN unifying perception, yield prediction & resource decisions: +23.7% yield accuracy (RMSE 0.87), +18.4% resource efficiency.
Bias-corrected LR (Monte-Carlo intervals + Sobol sensitivity) fused with a Random-Forest–CUSUM monitor; validated on Olympic-medal prediction.
Robust detection in fog & low-light via FAANet/PENet enhancement + GAM/BiFPN/SPD-conv; clear precision/recall/mAP gains (VanillaNet backbone best).
Full list on Google Scholar · ORCID
I'm open to AI / ML engineering and research opportunities. The fastest way to get a feel for my work is to ask my AI above — or just reach out.