🧑‍💻 AI Engineer • Agents / Computer Vision / NLP / Multimodal / Deep Search
Building grounded, explainable, and product-ready AI systems together.

• ⭐ 精选项目卡片
• 核心方向
• 代表项目（链接直达）
• 活动经历
• 联系方式

专注于将 Machine Learning / Deep Learning 能力工程化落地，核心方向包括：

• 🤖 Agents：深度搜索、不确定性感知、反思式迭代、Agent Collaboration、缓解RAG类检索增强的冲突与幻觉问题
• 💻 Computer Vision：视频修复与视频理解、超分辨率、OCR、医学影像分析
• 📝 NLP：文本理解与任务化生成、社媒舆论隐性恶意识别、情感分析
• 🎨 Multimodal：视觉-语言协同建模与跨模态推理, 建模图文特征关系
• 🔍 Deep Search 搜索算法：段落级检索、重排、推理链路与证据对齐
• 🏗️ 3D Reconstruction：ON Research real-world 3D reconstruction and digital twin applications

不仅关注模型在 benchmark 上的表现，更注重 可复现性、可解释性、以及生产环境的适配性。

Version 2.0 | 深度搜索 + Agents | 2026年2月

核心创新：

• 基于压力驱动（Pressure-Driven）的段落级迭代反思搜索架构，动态反思压力机制（Dynamic Reflection Stress）
• **五因子不确定性量化模型（5-Factor UQ Model）**自适应分配检索预算与推理深度
• 8种专业化角色Agent协作工作流，优化拆分搜索解析与高效协同
• 多样性去噪重排序 + NDCG/MRR 排序指标评估查询来源质量
• 异步高并发执行：asyncio + aiohttp 跨章节并发检索与调度

成果：

• ✨ 解决内容幻觉问题（算法量化确定性 88%⭡）
• 💴 Token成本优化（单次深度搜索约¥0.1）
• 🎯 强化推理稳定性、批判性、证据一致性

技术栈：Python | LLM Orchestration | asyncio | RAG | Uncertainty Quantification | MAS

Version 1.0 | OCR + 超分辨 + 文档结构化 | 2025年10月

问题：OCR 在税务票据、合同中低质扫描、拍照文档的识别准确率低

核心创新：

• Prior-Enhanced Attention 模块融入视觉+语义先验加权机制
• 从文本先验超分辨→文本检测→文本识别→文档结构化的端到端自动深度学习 pipeline
• 支持对图像扫描件的段落、字符、符号、字段标签的结构识别与组织，输出结构化JSON

成果：

• 📈 官方TextZoom资料源量化指标平均优化 5%⭡（PSNR, SSIM, Lrec, L1, CTCL）
• 🔝 鲁棒性增强：对模糊边缘与复杂背景的识别能力显著提升
• 🚀 下游自动化支持：企业表单处理、文档自动摘要等

技术栈：Python | PyTorch | Super Resolution | OCR | Computer Vision | Prior-Enhanced Attention

📄 Report Paper

Version 3.0 | 视频修复 + 深度学习 | 2025年

问题：SAM2 (Meta, 2024) 框架长时序列的记忆偏见积累问题；轻量化部署需求

核心创新：

• Conditional Memory Encoder + Hierarchical Selective Attention（HSA） 促进多层次语义理解
• 端到端可训练的语义驱动光流引导视频修复框架，对 E2FGVI (CVPR 2022) 创新优化
• NOF-Eraser 模块：仅利用轻量CNN特征提取 + Deformable Convolution + 光流辅助特征传播
• Mask-Agent 模块：文本语义解析自动识别用户关注区域，生成高质量掩码

成果：

• 📊 量化指标PSNR, SSIM, Sharpness, 伪影均优化 6%⭡ 原SAM-Based架构
• ⚡ 轻量部署：RTX 4070 Laptop 上实现 5.526 s/vsec (30fps), 12.839 s/vsec (60fps)
• 🎨 个性化效果：AI驱动掩码生成增强用户交互体验

技术栈：Python | PyTorch | Video Restoration | Optical Flow | Attention | Deformable Conv

📄 Report Paper

Team Competition Track | 多模态语义理解 + 隐性恶意识别 | 2025年9月至今

问题：社交媒体图文语义不一致、隐性指代、反讽式仇恨表达

核心创新：

• CLIP + LLMs 图像-文本联合框架，PE优化LLM输出
• 多个LLMs交叉验证的结果嵌入特征，增强社会先验建模（文化背景、网络热梗反讽）
• Cross-Modal Inference Graph：引入注意力矩阵显式建模图文特征关系
• GAT 编码跨模态图结构，实现多节点语义推理

成果：

• 🎯 多模态语义增强与隐性恶意识别系统
• 📈 高阶语义关联优化，缓解传统拼接式融合的信息损失
• 🔗 实体-语义-情感推导路径的显式刻画

技术栈：Python | CLIP | LLMs | Multimodal Fusion | GAT | Graph Neural Networks

Project | 机器学习 + 医学影像 | 2024年

问题：MRI脑肿瘤多分类中非典型小样本数据、病灶细粒度差异敏感性低

核心创新：

• 系统部署多种经典ML算法：SVM、XGBoost、Random Forest、KNN、LR、MLP
• 特征工程 + 传统模型组合策略提升小样本数据利用效率
• HOG + 统计纹理特征提取高判别性特征，减少大规模数据依赖
• 数据增强、特征归一化与降维方法有效缓解医学影像噪声与模糊

成果：

• 📊 性能指标：Acc=95.38%、F1=0.9553、Prec=96.07%、Rec=0.9551
• 💪 小样本学习：仅在1200份数据集达到优秀识别能力
• 🏥 临床适用：对少数病征判断的高精度识别能力

技术栈：Python | XGBoost | Random Forest | Feature Engineering | Medical Imaging

📄 Analysis Paper

• DataFun Conference 2025（北京 + 深圳）｜参会者｜2025年7月、11月
  参加 DataFunTalk 主办的 DACon 数智大会，系统化关注 Data + AI、大模型技术、数据智能与产业落地。重点学习大模型应用、搜索推荐系统、NLP、AI Agent、广告算法与数据架构，并了解 RAG、LLM 应用、ChatBI / Data Agent 等企业级方案。

• NVIDIA Technology Conference AI Summit（台湾）｜参会者｜2024年6月
  受邀参与 NVIDIA GTC 相关 AI 技术论坛与交流，重点关注 AI 在医学影像分析、三维重建、数字孪生等真实场景的应用；持续跟踪 VLM、GenAI、Robotics、Physical AI 等技术方向。

• 湾西小冰（中山）科技有限公司｜珠海市项目经理｜2024年
  负责珠海地区小冰数字人应用方案推广与优化，推进 AI 产品落地；面向客户需求提出定制化 C2C 方案，提升匹配度与满意度。

I focus on turning Machine Learning / Deep Learning into deployable systems, with strong interests in:

• Agents: uncertainty-aware planning, iterative reflection, tool orchestration
• Computer Vision: OCR, medical imaging, video restoration and understanding
• NLP: BERT-based modeling, language understanding and generation
• Multimodal: vision-language modeling and cross-modal reasoning
• Deep Search Algorithms: paragraph-level retrieval, reranking, evidence-aligned reasoning

I care not only about benchmark scores, but also reproducibility, explainability, and production readiness.

Version 2.0 | Agentic AI + Deep Search | Feb 2026

Core Innovations:

• Pressure-Driven Paragraph-Level Iterative Reflection Search Architecture with Dynamic Reflection Stress mechanism
• 5-Factor Uncertainty Quantification (5-Factor UQ Model) for adaptive retrieval budget & reasoning depth allocation
• 8 specialized role-based Agent collaboration workflows optimizing search decomposition & coordination
• Diversity-aware denoising reranking + NDCG/MRR metrics for source quality assessment
• Async high-concurrency execution: asyncio + aiohttp for cross-section parallel retrieval

Outcomes:

• ✨ Resolved content hallucination (88% algorithmic certainty⭡)
• 💴 Token cost optimization (~¥0.1 per deep search query)
• 🎯 Reinforced reasoning stability, criticality, evidence consistency

Tech Stack: Python | LLM Orchestration | asyncio | RAG | Uncertainty Quantification | MAS

Version 1.0 | OCR + Super-Resolution + Document Structuring | Internship Engineering | Oct 2025

Problem: Low OCR accuracy on low-quality document scans, receipts, contracts in enterprise scenarios

Core Innovations:

• Prior-Enhanced Attention Module incorporating visual + semantic prior weighting
• End-to-end automatic DL pipeline: text super-resolution → text detection → recognition → structure output
• Structure-aware parsing for paragraphs, characters, symbols, field labels; JSON-formatted structured output

Outcomes:

• 📈 Official TextZoom benchmark improvement: 5%⭡ across metrics (PSNR, SSIM, Lrec, L1, CTCL)
• 🔝 Enhanced robustness: improved recognition on blurred edges & complex backgrounds
• 🚀 Downstream automation support: enterprise form processing, automatic document summarization

Tech Stack: Python | PyTorch | Super Resolution | OCR | Computer Vision | Prior-Enhanced Attention

📄 Report Paper

Version 3.0 | Video Restoration + Deep Learning | 2025

Problem: SAM2 (Meta, 2024) memory bias accumulation in long sequences; lightweight deployment requirements

Core Innovations:

• Conditional Memory Encoder + Hierarchical Selective Attention (HSA) for multi-level semantic understanding
• End-to-end trainable semantic-driven optical-flow-guided video restoration framework; improved upon E2FGVI (CVPR 2022)
• NOF-Eraser module: lightweight CNN feature extraction + Deformable Convolution + optical-flow-assisted propagation
• Mask-Agent module: text semantic parsing automatically identifies user focus areas & generates high-quality masks

Outcomes:

• 📊 Metrics (PSNR, SSIM, Sharpness, Artifacts) improved 6%⭡ over SAM-Based baseline
• ⚡ Lightweight deployment: 5.526 s/vsec (30fps), 12.839 s/vsec (60fps) on RTX 4070 Laptop
• 🎨 Personalized effects: AI-driven mask generation enhances user interaction

Tech Stack: Python | PyTorch | Video Restoration | Optical Flow | Attention | Deformable Conv

📄 Report Paper

Team Competition | Multimodal Semantic Understanding + Implicit Hate Detection | Sep 2025 - Present

Problem: Social media image-text semantic mismatch, implicit references, ironic hate speech

Core Innovations:

• CLIP + LLMs joint image-text framework with PE optimization of LLM outputs
• Cross-verification results embedding from multiple LLMs enhancing social prior modeling (cultural context, memes)
• Cross-Modal Inference Graph: explicit attention matrix modeling image-text relationships
• GAT encoding of cross-modal graph structures enabling multi-node semantic reasoning

Outcomes:

• 🎯 Multimodal semantic enhancement & implicit hate detection system
• 📈 High-order semantic association optimization addressing fusion information loss
• 🔗 Explicit entity-semantic-emotion inference path modeling

Tech Stack: Python | CLIP | LLMs | Multimodal Fusion | GAT | Graph Neural Networks

Project | Machine Learning + Medical Imaging | 2024

Problem: Small-sample atypical data in MRI brain tumor multi-classification; low sensitivity to lesion fine-grained differences

Core Innovations:

• Systematic deployment of classic ML algorithms: SVM, XGBoost, Random Forest, KNN, LR, MLP
• Feature engineering + model combination strategy improving small-sample data efficiency
• HOG + statistical texture features for high-discriminative feature extraction, reducing large-scale data dependency
• Data augmentation, feature normalization & dimensionality reduction mitigating medical imaging noise

Outcomes:

• 📊 Performance metrics: Acc=95.38%, F1=0.9553, Prec=96.07%, Rec=0.9551
• 💪 Small-sample learning: excellent recognition on just 1200 samples
• 🏥 Clinical applicability: high-precision rare disease detection capability

Tech Stack: Python | XGBoost | Random Forest | Feature Engineering | Medical Imaging

📄 Analysis Paper

Repository | NLP Foundation Framework | Teaching & Practice

Content:

• Complete BERT pretraining & fine-tuning workflows
• Hugging Face ecosystem applications
• Modular NLP processing pipeline design
• Task implementations: text classification, named entity recognition, etc.

Tech Stack: Python | PyTorch | BERT | Hugging Face | NLP

• DataFun Conference 2025 (Beijing + Shenzhen) | Attendee | Jul & Nov 2025
  Participated in DACon by DataFunTalk with a structured focus on Data + AI, large models, data intelligence, and industrial deployment. Followed key topics including LLM applications, search/recommendation systems, NLP, AI Agents, ad algorithms, and data architecture, with practical exposure to RAG, ChatBI, and Data Agent solutions.

• NVIDIA Technology Conference AI Summit (Taiwan) | Attendee | Jun 2024
  Joined NVIDIA GTC-related AI technical sessions and exchanges, focusing on real-world AI applications in medical imaging, 3D reconstruction, and digital twins, while tracking VLM, GenAI, Robotics, and Physical AI trends.

• Wanxi Xiaobing (Zhongshan) Technology Co., Ltd. | Zhuhai Project Manager | 2024
  Led regional promotion and optimization of Xiaobing digital-human solutions in Zhuhai, with hands-on experience in AI product adoption. Worked directly with clients to propose customized C2C solutions and improve product-fit satisfaction.

• Modeling: ML/DL pipelines, transformer-style methods, multimodal fusion
• Agent Engineering: planning, reflection, retrieval augmentation, evaluation loops
• CV + NLP Integration: OCR, classification, video processing, language reasoning
• Implementation: Python, PyTorch ecosystem, experiment-driven iteration

• GitHub: https://github.com/SuleynanAuir
• Email Welcome to Connect 😊:
  • Daily suleynanaiur@gmail.com
  • Research & Academic t330034027@mail.uic.edu.cn
  • Domestic 2925795986@qq.com
• LinkedIn: https://www.linkedin.com/in/aiur-suleynan-1a58872b9/

Building AI systems that can search deeply, reason clearly, and deliver reliably.