HKUDS, LLMQuant, and X2Strategy: Toward a Personal Research and Quant Production OS
My recent work is converging around three systems:
- HKUDS ecosystem: AI agents, RAG, graph learning, recommender systems, research automation, and open-source research infrastructure.
- LLMQuant: a finance and quant-oriented project ecosystem with domain skills, market intelligence, strategy notes, and agent workflows.
- X2Strategy: a paper-to-strategy pipeline that turns research inputs into strategy specifications, code, backtests, and diagnosis reports.
I do not see these as separate learning tracks. I see them as three layers of a future personal research and quant production OS.
1
2
3
HKUDS -> AI / agent / RAG / research infrastructure
LLMQuant -> quantitative finance domain system
X2Strategy -> paper-to-strategy execution loop
Why These Three Fit Together
HKUDS is valuable because it shows how modern AI research systems are built: retrieval, agents, graph learning, recommender models, research automation, and evaluation.
LLMQuant is valuable because it brings the finance domain into the system: assets, markets, signals, events, risk, portfolio thinking, and analyst workflows.
X2Strategy is valuable because it focuses on execution. It tries to convert research material into a structured strategy spec, then into code, backtest artifacts, and diagnosis reports.
The fusion is natural:
1
2
3
4
5
6
7
8
research input
-> AI research infrastructure
-> finance domain framing
-> executable strategy specification
-> experiment / backtest
-> diagnosis
-> human review
-> reusable research artifact
My Integration Thesis
A useful AI research system should not stop at summarization.
It should help produce artifacts that can be reviewed, tested, improved, and reused:
- structured paper notes;
- knowledge cards;
- strategy specifications;
- experiment configs;
- baseline and ablation plans;
- backtest reports;
- failure analysis;
- draftable technical notes.
This is the difference between a chatbot and a research operating system.
1
2
chatbot = answer generation
research OS = artifact generation + validation loop + memory update
A Target Workflow
The workflow I want to build toward looks like this:
1
2
3
4
5
6
7
8
9
paper / market idea / research capture
-> personal capture system
-> HKUDS-style retrieval or agent analysis
-> LLMQuant-style domain framing
-> X2Strategy-style strategy specification
-> implementation or experiment plan
-> results and diagnosis
-> human approval
-> public-safe research artifact
This workflow keeps automation useful but bounded. The system can generate drafts, specs, experiments, and reports. Human review remains responsible for novelty, validity, risk, authorship, and final decisions.
Three Near-Term Research Lines
1. Research Automation
1
2
3
HKUDS AI-Researcher / DeepResearch-Eval
+ personal paper research workflow
-> paper-to-idea-to-experiment pipeline
This line is about turning reading into research output: hypotheses, experiment plans, results, and eventually draftable academic artifacts.
2. Quant Strategy Production
1
2
3
LLMQuant domain knowledge
+ X2Strategy strategy compiler
-> reproducible quant strategy experiments
This line is about turning financial papers, reports, and market ideas into structured strategy specs and testable backtest plans.
3. FICC Intelligence Workflow
1
2
3
HKUDS LightRAG / RAG-Anything
+ LLMQuant rates / FX / commodities framing
-> source-backed FICC analyst workflow
This line is about building AI workflows for real financial research tasks: rates, FX, commodities, macro signals, events, and decision support.
What I Want To Avoid
I do not want this to become a folder collection.
The goal is not:
1
download more repos
The goal is:
1
understand -> compare -> extract -> integrate -> verify
Every useful external project should be converted into a smaller artifact in my own system:
- a map;
- a note;
- a reusable schema;
- a workflow;
- a benchmark;
- a demo;
- a public-safe write-up.
First Practical Deliverable
The first deliverable should be a comparative note:
1
2
3
HKUDS AI-Researcher / DeepResearch-Eval
vs LLMQuant
vs X2Strategy
I want to compare them by:
- input format;
- output artifact;
- agent loop;
- evaluation method;
- strongest reusable component;
- how they connect to a paper research OS or a finance research workflow.
After that, the first pilot should be small:
1
2
3
4
one finance paper or market idea
-> domain-framed strategy spec
-> minimum experiment or backtest plan
-> diagnosis note
Long-Term Direction
The long-term goal is a personal research and quant production OS:
1
2
3
4
5
6
7
8
read papers
-> extract ideas
-> generate hypotheses
-> map to finance domains
-> produce strategy specs
-> run or plan experiments
-> evaluate results
-> write reusable research artifacts
This is the public-safe story I want to keep building: not an autonomous trading system, and not an automatic paper factory, but an auditable research production system that connects AI infrastructure, finance domain knowledge, and executable strategy research.