Quantitative trading research & infrastructure.
Quant research · LLM agent systems · AI infra · cloud native
My work spans two ends of the same problem. Above: quantitative research — factor mining, signal construction, strategy validation, live execution. Below: the infrastructure that holds it up — making the research loop run itself, run fast, and return results worth believing.
The current focus is rebuilding that research loop around LLM agents. The chain from reading a paper to a validated factor used to be strung together by hand; it now decomposes into agents that specialize — retrieval, hypothesis extraction, code generation, sandboxed execution, quality gating — and run end to end. The bottleneck in research was never a shortage of ideas; it was how fast you can disprove them. That is what the pipeline attacks.
On the engineering side I've built Go microservices, custom Kubernetes controllers, and service-mesh observability. That experience feeds straight back into AI infrastructure: concurrency orchestration, sandbox isolation, cost control and failure visibility in an agent system are still distributed systems problems — the unit of execution is just a model call instead of a process.
Quant research factor mining · signal construction · validation · execution
LLM agents multi-agent orchestration · codegen · sandboxing · gating
AI infra concurrency scheduling · cost control · observability
Cloud native Go microservices · Kubernetes controllers · service mesh
| autoalpha | Multi-agent factor mining pipeline. Takes a research direction, retrieves papers across arXiv/OpenAlex/Crossref, extracts factor hypotheses, generates and sandboxes the code, then gates on two-stage IS/OOS. Web console over FastAPI + SSE. Python FastAPI LLM agents |
| auto-trade | Automated order placement for the macOS build of a Chinese retail brokerage terminal, driven by the macOS Accessibility API — no Windows VM, no OCR, no simulated clicks. The native Cocoa UI reads back as a structured control tree, so state is verifiable instead of guessed. Python macOS AX automation |
| ICEEMDAN-TCN-FU | Time-series forecasting by ICEEMDAN decomposition, with a TCN fit per intrinsic mode. Python TCN |
| autopaper | Turns financial papers and research notes into backtestable strategies, then derives variants to build an IS/OOS-validated alpha library. Python LLM agents |
| stock-alpha | Full-stack alpha pipeline: minute-bar feature synthesis, IC-based factor selection, GRU + XGBoost hybrid for return prediction. Python PyTorch XGBoost |
| stock_prediction_cnn | Predicting stock direction with convolutional networks over price-series inputs. Python CNN |
Open to conversations about quant research, LLM agent systems, and AI infrastructure.
