Junxian Lyken Lin

Next.js 16 TypeScript FastAPI ReAct Agent Loop Native Tool Calling Gemma 4 31B (via Ollama) Neon PostgreSQL ChromaDB

多智能体数据副驾 (Cognitive Data IDE) Multi-Agent Data Copilot (Cognitive Data IDE)

扮演 AI 数据科学家的全自主 Web 数据 IDE A fully autonomous, Web-based Data IDE acting as an AI Data Scientist

基于前沿轻量级大模型（LLMs）构建，具备自主探索数据库、编写执行 Python/SQL 及渲染交互式图表的能力。采用双轨记忆系统 (SQLite + ChromaDB) 与后台异步压缩，提供高鲁棒性的 RAG 及多轮记忆交互。 Built around advanced lightweight LLMs, this agentic engine is capable of self-directed database schema exploration, native Python/SQL execution, and interactive chart rendering. It utilizes a dual memory system (SQLite & ChromaDB) with asynchronous compression for resilient RAG and multi-turn interaction.

核心亮点： Key Highlights:

智能体引擎：构建两步推理循环（探索架构 -> 运行代码），内置动态速率限制与防提示词注入（Recency Bias Fix）安全护栏。 Agentic Engine: Uses a two-step reasoning loop (Scan Schema -> Execute Code) with dynamic rate-limiting and prompt injection safeguards.
人机协同架构：采用三栏式“极客风”界面（历史记录 | 数据库架构 | 工作区），支持“直接执行”与“仅预览 SQL”的安全沙箱模式。 Human-in-the-Loop UI: Features a 3-column Hacker Theme layout with a safe dual-execution flow ("Execute Query" vs "Preview SQL Only").
量化评测与效率：在包含多重陷阱的严苛测试集上达到 75% 任务成功率 (6/8)，且查询的 Token 开销实现极致优化（约 1500 tokens/次）。 Evaluation & Efficiency: Achieved 75% Task Success Rate (6/8) on rigorous benchmark cases with highly optimized token efficiency (~1500 tokens/query).

运行应用 Launch App

Python LLM React FastAPI

智能食谱检索生成系统 (Agentic RecipeRAG) Agentic RecipeRAG

具备自主决策与量化评估框架的检索生成系统 RAG System with Autonomous Decision & Quantifiable Evaluation

设计并实现了一个基于 ReAct 架构的自主型 RAG (Agentic RAG) 系统，将传统的“硬编码逻辑路由”重构为“大模型原生工具调用”，并建立了一套自动化质量评估体系。 Designed and implemented an Agentic RAG system based on the ReAct architecture, refactoring traditional "hard-coded routing" into "native LLM tool-calling," along with an automated quality evaluation pipeline.

核心亮点： Key Highlights:

自主智能体：废弃静态路由，构建 ReAct 思考循环，实现“基于认知的自动化”。 Autonomous Agent: Discarded static routing, built a ReAct thought loop for cognitive-based automation.
生产级护栏：精细化 System Prompt 施加严格安全护栏，精准拦截无关请求。 Production Guardrails: Applied strict safety guardrails via precise prompt engineering to block irrelevant requests.
量化评测：开发自动化 Benchmark，引入 LLM-as-a-Judge 进行语义量化打分。 Quantifiable Eval: Developed an automated benchmark using the LLM-as-a-Judge paradigm for semantic scoring.

运行应用 Launch App

LLMs Statistical Scoring Dashboard

AI 驱动的 RFP 优化与工时管理 AI-Powered RFP Optimization & Capacity Management

Partner: CGI

分析劳动力规划工作流以识别高价值的流程低效点。设计一个结合大模型和统计评分的混合框架来自动化 RFP 工作量估算。 Analyzing workforce planning workflows to identify high-value process inefficiencies. Designing a hybrid framework using LLMs and statistical scoring to automate RFP effort estimation.

核心亮点： Key Highlights:

结合大模型与统计评分的混合 AI 框架。 Hybrid framework using LLMs and statistical scoring.
桥接技术信号与业务 ROI 指标的交互式仪表板。 Interactive dashboard bridging technical AI signals with business ROI metrics.
预计将为高管层挽回大量可计费工时。 Projected to recapture significant billable hours for senior leadership.

查看项目详情 View Details

Machine Learning PyTorch Python

视觉语言模型成员推断攻击 Vision-language Model Membership Inference Attack

Competition Rank #1 (Acc: 0.843)

在学校举办的视觉语言模型成员推断竞赛中，以 0.843 的测试集准确率（Accuracy）获得第一名。该项目探讨了多模态大模型在数据识别、隐私与安全性方面的潜在风险。 Achieved first place with a test set accuracy of 0.843 in a university-hosted Vision-Language Model membership inference competition. The project explores the risks of large multimodal models regarding data recognition, privacy, and security.

核心工作： Key Contributions:

特征工程：提取并分析了多模态模型的输出置信度与损失景观等行为信号，完成了特征降维。 Feature Engineering: Extracted and analyzed behavioral signals, such as output confidence and loss landscapes, to perform dimensionality reduction.
分类流水线：构建了二分类模型架构，并运用模型集成策略，有效提升了成员推断的准确率与泛化能力。 Classification Pipeline: Developed a binary classification architecture using ensembling strategies to improve inference accuracy and generalization.