Projects

Ten projects, from research to running system

Personal projects in quant research, forecasting, reinforcement learning and AI agents, alongside the shape of my work in banking and insurance. Filter by area, or arrive here from a skill to see where it's used.

Recorded run, step through it

AI trading swarm

A committee of AI analysts that reads point-in-time market data, cites every claim and hands its ideas to Strategy Orchestra to test, with a human approving each idea and every number reproduced before a strategy counts.

Watch one real idea go through the swarm: five analysts and a Critic on TQQQ, the thesis I approved, and the campaign that tested it, from 9,826 trades stopped at a gate to the Judge's 2.4 out of 100. Every line comes from the run's own records.

Python
PostgreSQL
TradingAgents
LangGraph
LM Studio
FastAPI
React

Read the write-up

royyang.ai/demos/ai-trading-swarm

RECORDED RUN · TQQQ

Data
Analysts
Critic
Thesis
Owner
Orchestra
Judge

DATAas of 24 Sep 2026 · 7 of 10 sources ready, 3 stated missing

AI agents and LLM tools

Agents, tools and LLM workflows where the model proposes and deterministic code decides.

01AI agents and LLM tools

Strategy Orchestra

Python
LangGraph
LLM agents
Tool use
pymoo

A multi-agent loop where an LLM proposes trading strategies and deterministic code tests, scores and routes them, with guards that stop the search from fooling itself across hundreds of trials.

strategy candidates proposed across 232 runs: 1,317; strategy candidates proposed across 232 runs
distinct strategies judged; the pass bar rises with every 50: 265; distinct strategies judged; the pass bar rises with every 50
tests pass, with 4 skipped by design: 692 / 696; tests pass, with 4 skipped by design

02AI agents and LLM tools

AI trading swarm

Python
PostgreSQL
TradingAgents
LangGraph
LM Studio

earlier strategies reproduced from stored data, 21 of 21 metrics each: 10 of 10; earlier strategies reproduced from stored data, 21 of 21 metrics each
tests pass on a clean copy, 281 of them against a real PostgreSQL: 808 of 824; tests pass on a clean copy, 281 of them against a real PostgreSQL
design decisions logged, each with its context and consequences: 71; design decisions logged, each with its context and consequences

03AI agents and LLM tools

Presence

Python
Ollama
SQLite
FastAPI
pydantic

A private conversational companion built from a WhatsApp chat history. It answers in one person's voice, grounded in what they actually wrote, and runs entirely on my own machine.

tests pass (1 skipped by design): 133; tests pass (1 skipped by design)
cloud model backends it can build: 0; cloud model backends it can build
LLM roles, each bounded by code: 5; LLM roles, each bounded by code

Forecasting and ML

Prediction on financial time series, evaluated against honest baselines.

04Forecasting and ML

TSBench: multi-series forecasting benchmark

Python
pandas
statsforecast
LightGBM
arch

A benchmark harness that runs forecasting models from several libraries over 16 years of 15-minute bars for 75 US tickers, scores them against finance-specific baselines, and picks winners with a Model Confidence Set.

15-minute bars, 75 tickers, 2010–2026: 7.2 M; 15-minute bars, 75 tickers, 2010–2026
known-answer tests pass on production data: 10 / 10; known-answer tests pass on production data
horizons where volume beats seasonal naive: 5 / 5; horizons where volume beats seasonal naive

Reinforcement learning

Sequential decisions under transaction costs.

05Reinforcement learning

Reinforcement-learning trading strategy

Python
stable-baselines3
Gymnasium
PyTorch
TA-Lib

Two agents, PPO and A2C, choose which of 10 low-correlation ETFs to trade every 15 minutes, and how much. The pipeline runs end to end. The evaluation doesn't hold up yet, and this page shows why.

ETFs kept for low correlation (mean ρ 0.12, against 0.42): 10 of 23; ETFs kept for low correlation (mean ρ 0.12, against 0.42)
inputs the agents read at every 15-minute step: 21,455; inputs the agents read at every 15-minute step
valid out-of-sample results so far: 4; valid out-of-sample results so far

Quant research

Point-in-time research on long histories of market data.

06Quant research

Portfolio Distiller

Python
pandas
scikit-learn
networkx
Jupyter

Cuts a universe of 75 US stocks and ETFs down to a 12-name basket without lookahead, comparing seven selection methods on one shared walk-forward harness.

names, chosen with data from before each decision only: 75 → 12; names, chosen with data from before each decision only
monthly out-of-sample rebalances, 2020–2026: 73; monthly out-of-sample rebalances, 2020–2026
max drawdown, best PCA method vs equal-weight: −12% vs −21%; max drawdown, best PCA method vs equal-weight

07Quant research

TA feature importance

Python
scikit-learn
TA-Lib
pandas
Ollama

Ranks about 2,500 technical-indicator columns by how well they predict 15-minute direction, without fooling itself, and uses a local LLM to port TradingView indicators behind deterministic causality checks.

indicator columns per split: ~2,500; indicator columns per split
LLM-translator self-test checks pass: 14 / 14; LLM-translator self-test checks pass
top features redundant (VIF > 10) before deduplication: 19 / 20; top features redundant (VIF > 10) before deduplication

08Quant researchIn progress

Algorithmic trading strategy guide

Python
Jupyter
MkDocs Material

A guide from strategy building to portfolio optimisation and risk management, with Jupyter notebooks to run alongside the text.

parts, from foundations to portfolio risk: 10; parts, from foundations to portfolio risk
chapters, plus 4 appendices: 25; chapters, plus 4 appendices

Trading systems

Engines with risk gates, reconciliation and recovery built in.

09Trading systems

Multi-strategy execution engine

Python
asyncio
SQLite
Interactive Brokers API
Docker

A local, asyncio-native engine that runs several strategies against Interactive Brokers, with risk checks, reconciliation and recovery designed in.

safety rules, each enforced in code and named in a test: 16; safety rules, each enforced in code and named in a test
tests pass with no broker connection, 37 of them fault injection: 1,033; tests pass with no broker connection, 37 of them fault injection
order cycles on a paper account; the first five found an order-ID bug: 8; order cycles on a paper account; the first five found an order-ID bug

Infrastructure and tooling

The local AI stack and workflows I build with.

10Infrastructure and tooling

Home AI platform

Ollama
LM Studio
TypeScript
pgvector
Cloudflare Tunnel

An invite-only chat platform that serves local LLMs from a Mac at home, with one gateway that checks every request and keeps private documents away from cloud models.

tests pass; the API's 510 run against a real Postgres and Redis: 588; tests pass; the API's 510 run against a real Postgres and Redis
exposure checks pass on the public tunnel, with admin closed at the edge: 6 of 6; exposure checks pass on the public tunnel, with admin closed at the edge
unified memory, M4: 128 GB; unified memory, M4

Want the full detail, with dates and stack?

View my résumé