研究与探索 Research & Explorations
正在进行
Google TimesFM LoRA
状态： 评估中 · 非公开
方向： 时间序列基础模型、参数高效微调、金融风险控制
本项目探索使用 AI Agent 量化金融投研系统积累的自有历史金融数据，对 Google TimesFM 进行微调，评估模型在金融时间序列预测任务中的泛化能力。
该模型计划作为决策辅助风险控制系统的组成部分，为相关风险判断提供时间序列预测信息。目前项目处于方案与可行性评估阶段，数据内容、训练方法和实验结果暂不公开。
AI Agent 量化金融投研系统
状态： 进行中 · 部分非公开
方向： AI Agent、量化研究、执行审计、系统自迭代、数据治理
本项目探索如何将 AI Agent 应用于持续运行的量化金融投研流程，并在复杂业务链路中保持执行过程的可追溯性、可评估性与可控性。
当前关注的问题包括 Agent 工作流与 Prompt Engineering、Prompt／策略／模型的多版本管理、结构化执行审计、自动 Review 与影子实验，以及数据质量、运行状态和 LLM 成本的持续观测。系统采用 AI 深度参与开发、人工负责关键 Review 与上线决策的协作方式，并通过持续迭代验证复杂 AI 系统的可维护性和演进能力。
该项目同时为 TimesFM 微调研究积累历史金融数据，并计划将时间序列预测能力接入决策辅助风险控制流程。具体业务逻辑、数据内容、策略和系统实现细节暂不公开。
已结束
CAMINO 云原生自主管理与意图编排器
状态： 阶段完成
方向： 云原生、GitOps、多边缘编排、资源预测、可观测性
相关内容： 硕士毕业设计研究记录
本项目围绕多边缘环境中的意图驱动服务编排展开，探索如何通过云原生与 GitOps 方法实现分布式边缘服务的自动部署、监控和调度。
研究实现了 CAMINO 控制平面的主要能力，并提出由意图分析与资源预测组成的 NOMRPS 方案。模拟实验表明，与基线部署方式相比，该方案能够降低因单节点资源不足造成的部署失败风险；服务间平均通信延迟降低约 25%，P90 延迟降低约 55%。分布式部署提高了 CPU 资源利用效率，但平均内存使用量增加了 51.48%。
受项目周期限制，资源预测部分尚未完成真实模型训练，相关实验主要验证意图拆解与编排策略，而非完整的在线预测能力。这一限制也构成了后续研究和改进方向。
基于 RAG 的 LLM 文案生成系统
状态： 实验结束
方向： RAG、Few-Shot Prompting、风格一致性
本项目探索如何改善 LLM 生成文案时风格不稳定、不统一的问题。
系统通过 RAG 检索与目标风格相关的历史文案，并将检索结果作为 Few-Shot 示例提供给 LLM，以建立相对稳定的风格锚点。实验最终验证了该方法的基本可用性，生成结果能够在一定程度上保持目标文案风格。
项目完成可用性验证后未继续打磨或部署上线，目前没有可公开的代码、数据或实验成果。
基于 LLM 的金融市场情绪分析与决策系统
状态： 已停止
方向： 金融新闻分析、市场情绪、LLM 信息抽取
本项目探索使用 LLM 从多来源金融新闻中提取市场情绪及其关联板块，并据此计算不同板块的整体情绪指标，为市场决策提供辅助信息。
系统定时获取多个新闻数据源，由 LLM 解析新闻内容、判断情绪倾向并识别相关行业板块，再对结果进行聚合。
在实际评估中，项目发现新闻信息相对于股票市场价格变化具有天然滞后性，而基于新闻进一步计算的情绪信号延迟更加明显。相关数据未能表现出足够高的决策价值，因此项目停止继续开发。
该项目没有可公开的代码、数据或实验成果，但其结论为后续量化金融系统的数据选择和信号评估提供了经验。
调研与早期探索
OpenDayLight 与 SDN 平台调研
类型： 前期技术调研
方向： SDN、网络控制、分布式网络
该调研源于硕士毕业设计早期的方向探索，主要关注 OpenDayLight 及相关 SDN 技术在集中式虚拟网络与网络管理场景中的应用可能性。
随着毕业设计研究方向调整，该路线没有继续发展为正式项目。相关内容作为技术选型与研究方向演变的记录保留。
In Progress
Google TimesFM LoRA
Status: Under Evaluation · Non-public
Topics: Time-Series Foundation Models, Parameter-Efficient Fine-Tuning, Financial Risk Control
This project explores fine-tuning Google TimesFM on proprietary historical financial data accumulated by the AI-agent-based quantitative research system. Its primary objective is to evaluate the model’s generalisation capability in financial time-series forecasting.
The resulting model is intended to serve as one component of a decision-support risk-control system by providing time-series forecasts for related risk assessments. The project is currently undergoing feasibility and methodology evaluation. Details of the dataset, training process, and experimental results are not publicly available.
AI-Agent-Based Quantitative Financial Research System
Status: In Progress · Partially Non-public
Topics: AI Agents, Quantitative Research, Execution Auditing, System Self-Improvement, Data Governance
This project explores how AI agents can support a continuously operating quantitative financial research workflow while keeping execution traceable, assessable, and controllable across complex business processes.
Current areas of focus include agent workflow design and prompt engineering, version management for prompts, strategies, and models, structured execution auditing, automated reviews and shadow experiments, and continuous observation of data quality, system health, and LLM costs. The system is developed through a collaborative approach in which AI participates extensively in implementation while humans retain responsibility for critical reviews and deployment decisions. Continuous iteration is used to evaluate the maintainability and evolvability of a complex AI system.
The project also provides historical financial data for the TimesFM fine-tuning study, with time-series forecasting planned as part of a decision-support risk-control workflow. Specific business logic, datasets, strategies, and implementation details are not publicly available.
Concluded
CAMINO: Cloud-Native Autonomous Management and Intent-Based Orchestrator
Status: Milestone Completed
Topics: Cloud-Native Systems, GitOps, Multi-Edge Orchestration, Resource Prediction, Observability
Related Content: MSc Project Research Notes
This project investigated intent-driven service orchestration across multi-edge environments, exploring how cloud-native and GitOps practices can support automated deployment, monitoring, and scheduling of distributed edge services.
The research implemented the main capabilities of the CAMINO control plane and proposed NOMRPS, a design combining intent analysis with resource prediction. In simulation, the approach reduced deployment failures caused by single-node resource constraints. Compared with the baseline deployment method, average inter-service communication latency decreased by approximately 25%, while P90 latency decreased by approximately 55%. Distributed deployment improved CPU resource utilisation but increased average memory usage by 51.48%.
Due to the limited project timeframe, the resource-prediction component was not trained with a production model. The experiments therefore primarily validated the intent decomposition and orchestration strategy rather than a complete online prediction capability. This limitation also defines a clear direction for future work.
RAG-Based LLM Copywriting System
Status: Experiment Concluded
Topics: Retrieval-Augmented Generation, Few-Shot Prompting, Style Consistency
This project explored how to improve stylistic consistency in copy generated by large language models.
The system used retrieval-augmented generation to retrieve historical copy relevant to a target style. The retrieved examples were then provided to the LLM as few-shot demonstrations, creating a more stable stylistic anchor during generation. The experiment established the basic feasibility of the approach, with generated content showing a reasonable degree of consistency with the intended style.
The project was not further refined or deployed after the initial feasibility validation. No source code, dataset, or experimental results are currently available for public release.
LLM-Based Financial Market Sentiment Analysis and Decision-Support System
Status: Discontinued
Topics: Financial News Analysis, Market Sentiment, LLM-Based Information Extraction
This project explored the use of large language models to extract market sentiment and associated industry sectors from financial news collected across multiple sources. The extracted signals were aggregated into sector-level sentiment indicators intended to support market analysis and decision-making.
The system periodically collected news from several data sources. An LLM analysed each item to determine its sentiment and identify the market sectors to which it was relevant, after which the results were aggregated.
Evaluation revealed a fundamental limitation: news sources naturally lag behind movements in stock prices, while sentiment signals derived from that news introduce further delay. The resulting data did not demonstrate sufficient value for decision support, so further development was discontinued.
Although no code, data, or experimental results are available for public release, the negative finding informed subsequent decisions about data selection and signal evaluation in later quantitative-finance work.
Preliminary Studies and Early Explorations
OpenDayLight and SDN Platform Study
Type: Preliminary Technical Study
Topics: Software-Defined Networking, Network Control, Distributed Networks
This study formed part of the early exploration for my MSc project. It examined OpenDayLight and related software-defined networking technologies, with particular interest in their potential use in centrally managed virtual networks and network-management systems.
The direction was not developed into a full project after the MSc research topic changed. It is retained as a record of the technical evaluation and the evolution of the project’s research direction.