Project
Atlas
AI that trades for you — on your terms.
What it is
Multi-agent trading system · live paper trading on Alpaca · final-year capstone
By the numbers
- 27+
- API routes
- 9 nodes
- Agent graph
- 7
- Inngest functions
- Supabase + Mongo
- Databases
Tech stack
Atlas — Agentic AI Trading System
Capstone project, SIT | B2C SaaS | Solo build
Atlas is my academic capstone and a commercial product in one. The question it's built around: how much should AI be trusted to act autonomously with real money?
The Core Idea — Execution Boundary Controller
The Execution Boundary Controller (EBC) is the architectural centrepiece. It gives users a configurable slider between two modes:
- Advisory — AI analyses and recommends. Human approves every trade.
- Autonomous — AI executes at ≥65% confidence. A 5-minute override window lets the user intervene before the order fires.
Same LangGraph pipeline runs either way. The mode determines what happens at the execution boundary.
The Agent Graph
A multi-agent system built on LangGraph with parallel analyst nodes — technical, fundamental, and sentiment analysis running concurrently — fanning into a synthesis agent, then risk assessment, then portfolio decision. Backed by Gemini 2.5 Flash with structured JSON outputs throughout.
Backtesting Engine
The backtesting engine replays the real AI pipeline across historical dates with strict look-ahead bias prevention. It computes Sharpe ratio, CAGR, Calmar ratio, max drawdown, and profit factor — giving users a truthful read on how the strategy would have performed.
Stack
Next.js 16, LangGraph.js, Gemini 2.5 Flash, Inngest, Supabase, MongoDB Atlas, Clerk, Alpaca