Projects
AlphaBot — Multi-LLM Agent System for Systematic Alpha Discovery
Quantitative Research Intern, Cubist (Point72), Sep 2025 – Feb 2026; independent self-funded deployment ongoing.
A closed-loop multi-LLM pipeline for systematic alpha discovery, deployed and validated on two markets.
- Multi-LLM proposal, harness-gated evaluation. Three frontier model families generate factor expressions in parallel; a backtesting harness — calibrated to hedge-fund-grade quant-research review (rolling-period statistical validation, odd/even sample consistency, joint train/validate thresholds, strict out-of-sample holdout) — is what decides which expressions survive. The language models propose; the harness disposes.
- Pipeline. Center-of-expertise generation → multi-LLM brainstorm → consolidate / parse / repair → backtest with train-validate splits → parameter tuning on survivors → out-of-sample holdout.
- Two-market validation.
- US equities at Cubist. 80+ significant meta-alphas across momentum, mean reversion, liquidity, information flow and other risk factors; ensemble prediction models attain promising Sharpe on both large- and small-cap US equities. (Specifics IP-protected; qualitative summary only.)
- Self-funded crypto. An independent deployment on mid-frequency crypto outside Cubist, joint with Beier Liu — public live performance over 18 months.
- A research finding from the multi-LLM work. Across the three frontier model families used in the ensemble, research-output quality on this task ranks Opus > GPT > Gemini.
A sanitized technical writeup is in this blog post; deeper details pending Cubist review.
IvorySquare
An open Ivory Tower for everyone. Apr 2026 – present. In collaboration with Han Yan.
A framework that treats peer-reviewed methodology — across finance, accounting, economics, and operations research — as a first-class tool surface for LLM agents. Skills are paper-derived, citation-grounded, and gated by purpose-built evaluation harnesses; the human-expert layer remains disjoint from engineering through declarative persona contracts.
- Two-tier skill graph. A foundational concept layer at textbook-subsection granularity, sitting under prerequisite ordering, and a paper-derived methodology layer with formal implementations, worked-example replication, and line-item citation provenance back to the source filing.
- Tool surfaces. Every skill — foundational and paper-derived — exposes a declarative interface callable from either an MCP server or an OpenAI tool specification. One library, two surfaces, no translation glue.
- Persona-driven authorship. Four LLM persona configurations — accounting expert, quant-finance methodologist, evaluation agent, citation auditor — author and gate every artifact, keeping domain expertise as YAML/markdown rather than Python.
- Motivating research direction. Academic citation networks as a structured post-training substrate for tool-using LLMs: each skill supplies both a tool-use trace and a verifiable ground-truth signal, and the citation topology gives a natural curriculum from primitive methods to composite ones.
- Open source. github.com/SenYangOM/IvorySquareSolutions.
A sanitized architectural deep dive is in this blog post.