I enjoy turning ambitious ideas into systems that actually runβfrom model experiments and multi-agent workflows to backend services, deployment infrastructure and user-facing products.
I believe future software will increasingly be operated through AI. Instead of learning every complicated interface, users will describe their goals, while agents select tools, execute workflows and iteratively improve the results.
AlphaPilot is an open-source, LLM-driven quantitative research and trading platform.
It connects the complete workflow from a financial hypothesis to a validated strategy:
- Multi-agent alpha factor discovery
- LLM, genetic programming and reinforcement learning-based factor generation
- Qlib-powered factor evaluation and portfolio backtesting
- Factor and strategy asset management
- Quantitative timing and daily trading signals
- Web Portal, CLI and remote task notifications
- Controlled paper, simulation and live-trading infrastructure
- Extensible data, strategy, portfolio and broker plugins
The long-term goal is to make AlphaPilot a professional capability layer for AI agents: users describe a research objective in natural language, while the agent queries data, generates factors, runs experiments, analyzes failures and continuously improves the strategy.
I am working on exposing AlphaPilot through standardized APIs and Agent Skills.
In this architecture:
User
β Natural Language
Local AI Agent
β Planning Β· Memory Β· Tool Selection
AlphaPilot Skills / API
β
Data β Factor Mining β Backtesting β Strategy β Paper Trading
The AI agent acts as the main interaction layer, while AlphaPilot runs as an independent and auditable quantitative engine.
I am also exploring how reinforcement learning can improve agents that interact with tools and real environments, with a focus on:
- Multi-turn tool-use training
- Asynchronous rollout and distributed sampling
- Reward design and verifiable feedback
- PPO, GRPO, RLOO and REINFORCE++
- Efficient training with vLLM, Ray and DeepSpeed
- Agent evaluation, reliability and self-improvement
| Project | Description |
|---|---|
| AlphaPilot | LLM-driven quantitative research, backtesting and controlled trading platform |
| AlphaPilotArena | Mobile-first public testing platform for factor hypotheses and standardized backtests |
| AlphaPilotArenaWorker | Isolated worker and runner infrastructure for quantitative research tasks |
| AgenticRL_P | Experiments and practice around Agentic Reinforcement Learning |
| GFPCC | Federated learning and graph-based proactive caching research |
Agentic RL LLM Agents & Tool Use
LLM Systems AI Infrastructure
Distributed Training High-Performance Inference
AI Γ Quant Multi-Agent Collaboration
Reliable AI Systems Open-Source Engineering
I am interested in collaborating on:
- Agentic RL and model post-training
- AI agents and agent infrastructure
- LLM-powered developer tools
- Quantitative research systems
- High-performance AI serving
- Open-source projects with real users
- Ambitious and slightly crazy hackathon ideas
For me, open source is more than publishing code. It is a way to document ideas, make experiments reproducible, learn in public and build things that can continue growing beyond a single developer.
Think boldly. Build quickly. Turn tokens into tools.