Academic Homepage · Google Scholar · LinkedIn · CV
I am a Data Science undergraduate at HKUST(GZ) and a quantitative research intern. My primary research and career axis is Financial ML / AI-related Quant, with a broader interest in reliable learning from non-stationary sequential data and in AI systems whose state, evidence, time, and failure boundaries are explicit.
This GitHub profile is my engineering and public-research evidence surface. For publications, research trajectory, education, PhD-facing research direction, and the canonical CV, use my academic homepage.
| Axis | Questions I care about | Current evidence |
|---|---|---|
| Financial ML / Quantitative Research | How do models remain valid under non-stationarity, changing information sets, signal redundancy, and realistic temporal evaluation? | LENS (ACM ICAIF 2025), real-market Level-2 research, AlphaSeeker, quantitative internship |
| Reliable AI & Decision Systems | How should high-stakes systems expose temporal provenance, policy boundaries, uncertainty, and decision lineage instead of hiding them behind model output? | Clausula, Reliable Multi-Agent Financial Forecasting FYP |
| Agent / Research Systems | How can autonomous research and tool-using systems make state, provider semantics, evidence, and failure behavior auditable? | DSHelm, upstream Senpi / oh-my-openagent work, Windows computer-use bridge |
My 2027 PhD applications are centered on the methodological overlap between Financial ML, non-stationary sequential learning, robust evaluation, reliable AI, and data/agent systems. Long term, I expect to work close to quantitative research and Financial AI while keeping the underlying research questions transferable beyond a single market or model family.
| Project | Role | Why it matters to my profile |
|---|---|---|
| Clausula | creator / maintainer | Local-first deterministic investment decision system. Versioned ledgers, point-in-time provenance, policies, research evidence, capital/risk boundaries, and decision memory remain canonical outside the LLM. This is the clearest bridge between my Financial ML interests and reliable-system design. |
| DSHelm | creator / maintainer | Explainable multi-model routing for DeepSeek Harness with evidence-backed policy resolution, compatibility boundaries, Resolution Trace, and reproducible execution fixtures. |
| AlphaSeeker-TradeMaster | research artifact maintainer | Financial time-series forecasting artifact from TradeMaster Cup 2026 with chronological validation and an explicit audit of offline-versus-causal information boundaries. |
| Project | Surface | Engineering signal |
|---|---|---|
| LedgerNest · EN overview | self-hosted collaborative accounting | explicit accounting semantics, multi-user isolation, audit trails, import/export correctness, mobile-first product delivery |
| Keji · EN overview | self-hosted client-work CRM | privacy-oriented vertical product engineering, document workflows, permissions, backup/restore, operational auditability |
| dsh-computer-use-windows | Windows computer-use bridge | OCR-grounded actions, bounded verification, explicit failure evidence, hosted-CI versus real-desktop validation boundaries |
Small personal and experimental repositories remain public when they have a useful independent boundary, but they are intentionally not part of the main research / career narrative.
| Area | Contribution | Upstream outcome |
|---|---|---|
| Session isolation | oh-my-openagent #6829: diagnosed cross-session ULW continuation state leaking between independent sessions sharing one working directory; proposed explicit session scoping and fail-closed status reads. | Maintainer called the diagnosis “correct and load-bearing”; the final upstream fix adopted the session-scope module boundary and credited me through co-authorship. |
| SDK failure semantics | senpi #1223: centralized terminal-result failure classification across streaming, managed failover, resident settlement, and successful-turn bookkeeping. | The shared failure-classification design was adopted and explicitly credited in the merged upstream implementation. |
| Provider/cache semantics | senpi #1237: models prompt-cache lifetime as fixed, automatic, disabled, or unknown, avoiding fabricated TTL/savings claims for provider-managed caching. |
Active upstream review. |
| Release correctness | senpi #1224: preserves the Linux x64 PTY prebuild through build → artifact staging → npm packaging and makes required-target publication fail closed. | Active upstream review. |
reproduce → identify the invariant → locate the ownership boundary → make time/state explicit → add regression evidence → implement the smallest durable fix → state residual limits
Across Financial ML and systems work, I repeatedly care about the same failure mode: a result can look correct while silently depending on information, state, credentials, execution conditions, or assumptions that will not hold later. My research and engineering both try to surface those dependencies before they become hidden sources of error.
| Surface | Canonical role |
|---|---|
| Academic homepage | PhD / research-job visual CV: research focus, future direction, publications, experience, education, distinctions |
| GitHub profile | engineering depth, public research artifacts, project ownership, upstream review outcomes |
| Google Scholar | publication record |
| professional trajectory and external career identity |
Public repositories are intentionally separated from proprietary market data, employer IP, credentials, personal financial records, and private research assets.