Real-time AI trading
Le Flux
A multi-model trading arena that turns live crowd positioning and market data into scheduled AI decisions.
Professional product
- 01Gemini FlashLong
- 02Llama MaverickFlat
- 03DeepSeekShort
Privacy-safe interface composition based on the shipped product.
My contribution
Built the complete arena platform: TypeScript API, PostgreSQL model, real-time services, LLM orchestration, dashboard, monitor, and deployment.
The challenge
What the system had to solve
Live market data, crowd behavior, multiple AI models, cost controls, and trading-session rules had to stay synchronized in one observable system.
Product surface
Modules and functionality
- 01
AI model seats and scheduled cycles
- 02
Live crowd-position ingestion
- 03
Market bars and session controls
- 04
Leaderboard, trades, and P&L history
- 05
Prompt registry and cost tracking
- 06
Live risk monitor and WebSocket updates
Engineering
Decisions that shaped the build
- Orchestrated multiple OpenRouter models from one controlled scheduler.
- Combined position polling, market-data WebSockets, bar aggregation, and PostgreSQL persistence.
- Kept long-lived API credentials out of the browser through a server-side PHP proxy.
- Preserved an existing monitor contract while adding the new Arena API surface.
Outcome
What the work delivered
- Created one live environment for comparing model decisions against the same market context.
- Made model behavior, costs, trades, and rankings observable from a dedicated dashboard.
- Deployed independent server and dashboard surfaces with automated restart and delivery workflows.
Technology
Core stack
- TypeScript
- Fastify
- PostgreSQL
- OpenRouter
- WebSockets
- PHP