Nanjing University · M.S. 2027 (expected)
I work across LLMs, agents, multimodal learning, and AI for Science, from model training and inference to evaluation and applications. My experience includes AI algorithm internships at Huawei and Alibaba.
Competitions · Experience · Research · Contact
|
🥈 Silver · 58 / 4,186 teams · Top 1.39% Co-hosted by OpenAI, Google and IEEE. An international challenge on reproducible failures in tool-using agents. Model-specific search, tool-trace feedback, and replayable security evaluation.
|
🥈 Silver · 42 / 2,445 teams · Top 1.72% Hosted by Google Jigsaw / Conversation AI. A Kaggle Featured competition on real-world community moderation. Rule-conditioned LLM judging, efficient post-training, inference, and model ensembles.
|
|
🥈 Silver · 83 / 3,803 teams · Top 2.18% Led by CSIRO, Australia's national science agency, with Google Australia and Meat & Livestock Australia. Dual-view visual features and structured prediction of five pasture biomass targets.
|
🥉 Bronze · 217 / 2,240 teams · Top 9.69% NeurIPS 2025 official Competition Track. Hosted by the University of Notre Dame. Molecular representations and complementary models for material property prediction.
|
B leaderboard: 7th · National finalist
A challenge founded by Tsinghua University's BNRist and the Tsinghua–Tencent Joint Laboratory, under the guidance and funding of NSFC's Department of Information Sciences. Official event background
Fine-grained BI-RADS grading with multi-scale features and ordinal learning.
Explore competition methods and implementation details
Team Silver Medal · Jun–Sep 2026 · Official competition
- Contributed to 40 configurations across GPT-OSS and Gemma, evaluating actual tool execution through traces.
- Configured 30 probes per strategy, ranking strategies by a heuristic estimate of tool-action reward per second over the final 20 probes.
- Supported up to 2,000 candidates per model and a single-file SDK submission for independent replay; analyzed public/hidden guardrail differences and removed dependence on a fixed test marker.
Silver Medal · Top 1.72%
- Fine-tuned Qwen3-4B with 4-bit QLoRA, completion-only supervision, and single-token scoring for rule-conditioned moderation.
- Combined generative LLM, GTE + Triplet Loss, and DeBERTa-v3 branches through rank blending.
- Used DeepSpeed ZeRO-2 for efficient training and vLLM for batch inference, including generalization to unseen rules.
- Used 357 high-resolution images to predict five biomass targets with a shared DINOv3 ViT-Huge backbone and a Gated Local Token Mixer.
- Predicted Green, Dead, and Clover biomass and reconstructed GDM and Total to preserve target relationships.
- Added SigLIP + LightGBM / CatBoost, four-fold validation, and late blending; improved weighted R² from 0.54 to 0.63.
- Combined RDKit / Mordred descriptors, molecular fingerprints, graph models, and gradient-boosted trees.
- Addressed missing labels, different target scales, feature selection, and multi-target validation.
- Combined EfficientNetV2, multi-scale features, BN-Neck, and Multi-Dropout for limited and imbalanced data.
- Explored cumulative ordinal labels with BCEWithLogitsLoss; used five-fold validation, TTA, checkpoint ensembles, and Jittor AMP.
China Merchants Bank Digital Finance Bootcamp · Jun 2026
Training track: individual 1st place · Team Silver Award
Built a Qwen3.5-2B training and execution-verification workflow across 4 database schemas and 5,635 interaction turns. Improved execution-verified SQL labels from 36.6% to 97.32%.
Post-training and verification details
- Applied LoRA, SFT, DPO, and GRPO / RLVR, using real query execution feedback for supervision and preference construction.
- Corrected 345 timezone-related false errors; validated 5,484 / 5,635 SQL labels.
- Built a PostgreSQL execution verifier for execution status, result consistency, and error analysis.
- Training framework: Twinkle.
Jul 2026–Present · Image Processing Project Lead · Outstanding Intern
Developed Retarget-Agent for content-aware image adaptation, preserving people, products, text, and logos.
Evaluated 600+ business images across 8 routes. A-grade rates on the fixed evaluation set:
| Best single method | Agent selection | + Generative repair |
|---|---|---|
| 77.33% | 93.00% | 96.33% |
Responsibilities and technical details
- Established a quality rubric covering content fidelity, visual completeness, and composition.
- Combined detection, OCR, segmentation, and geometric processing with multimodal judging to generate, score, and select candidates.
- Added controlled generative repair with original subject/text restoration, re-evaluation, and fallback.
- Delivered FastAPI interfaces and replay-based evaluation, including versioned results and failure analysis.
Stack: Python · OpenCV · YOLO · OCR · SAM · ComfyUI · FLUX · FastAPI
Nov 2025–Feb 2026 · Virtual Try-on
Developed AI-generated product imagery for apparel merchants, focusing on garment detail, identity consistency, and production efficiency.
500+ real product/model images · Selected merchant assets: ~7 days of traditional shooting → ~30 minutes
Responsibilities and technical details
- Built controllable try-on workflows with FLUX.1 Fill + Redux and Gemini image editing, combining garment references, model images, masks, and prompts.
- Developed preprocessing, automatic captioning, workflow-to-API conversion, batch inference, and configurable FastAPI interfaces.
- Fine-tuned and integrated Face-Fidelity LoRA, covering data preparation, captioning, inference tuning, and dynamic ComfyUI loading.
- Designed quality-based routing and review for garment texture, faces, and occlusion boundaries, with bounded retries for failed outputs.
Stack: Python · FLUX · Gemini · LoRA · SigLIP · ComfyUI · FastAPI
First-author work in molecular machine learning and scientific computing.
| Research | Selected results |
|---|---|
| ⚛️ PRISM · Multi-fidelity Equivariant Neural Potentials | 200 MP2 configurations · Energy MAE 0.48 meV/atom · Force MAE 42.31 meV/Å |
| 🧪 Hydrogen Transport at Mineral Interfaces | 40k+ atomic configurations · DFT → ML potentials → molecular dynamics |
Research methods and benchmark details
PRISM
- Learned medium- and high-fidelity energy/force corrections to a low-cost physical baseline using a shared equivariant graph representation.
- Combined multi-fidelity residual learning, joint energy–force training, data alignment, and controlled ablations.
- On PBC-Water, reduced energy error by 67.3% vs. Δ-EGNN and force error by 26.4% vs. Δ-PaiNN in the reported benchmark comparisons.
- Stack: PyTorch · SchNetPack · Equivariant GNNs · Linux / HPC.
Hydrogen transport · Jiangsu Provincial Frontier Technology R&D Project
- Connected structure sampling, VASP labeling, data quality checks, DeepMD / DPA / MACE training, and LAMMPS simulations.
- Analyzed aggregation and transport in nanoconfined brine and mineral interfaces through density, clustering, diffusion, orientation, and residence-time statistics.
| Build & deploy | Python · C/C++ · Linux · Git · FastAPI · SQLite · PostgreSQL |
| Train & evaluate | PyTorch · Jittor · Qwen · LoRA / QLoRA · SFT / DPO / GRPO · DeepSpeed · vLLM |
| Models & applications | LangGraph · AgentScope · OpenCV · DINOv3 · SigLIP · FLUX · ComfyUI · LightGBM / CatBoost |
| Scientific computing | SchNetPack · RDKit · ASE · VASP · DeepMD · MACE · LAMMPS |
Nanjing University · M.S. in Physical Chemistry · 2024–2027 (expected)
Kuang Yaming Honors School · AI for Science
Email ·
All repositories







