AI-Powered Smart Trading & Portfolio Platform — built for the III 5.0 Hackathon.
TradeWise is a full-stack paper-trading platform that combines real-time market data, ML-ensemble price predictions, autonomous portfolio management, and natural-language trade execution in a single polished interface.
cp backend/.env.example backend/.env
# Fill in API keys (see Environment Variables below)
docker compose up --buildVisit http://localhost:3000 and log in with demo@tradewise.cloud / demo123.
- Paper Portfolio — buy/sell any US equity with $100k starting balance; live P&L, realized gains, transaction history
- Watchlist — track symbols with live price ticks via Finnhub WebSocket
- Real-time Prices — WebSocket price feed with simulated random-walk when markets are closed
- ML Ensemble Predictions — XGBoost + Logistic Regression + LSTM trained on 1Y OHLCV; majority vote direction with confidence score (capped 85%)
- FinBERT Sentiment — HuggingFace FinBERT cascade with Gemini Flash fallback, fused into risk scoring
- AI Advisor — ReAct agent (Gemini 2.5 Flash) with tool use: portfolio lookup, news fetch, price quotes, technicals; SSE-streamed responses
- Multi-Agent Swarm — parallel specialist agents (momentum, mean-reversion, sentiment, risk) that vote on trade recommendations
- Risk Analysis — Monte Carlo simulations for VaR, CVaR, drawdown projections
- Autopilot — autonomous portfolio manager; configurable modes (Conservative → Full Auto) with guardrails (max drawdown, daily loss, position size limits); real-time thought stream via SSE
- Signal Scanner — background scanner across 65-symbol universe every 5 minutes; 6 signal types (ML Consensus, Breakout, Oversold, Overbought, Momentum, Reversal) with 4-indicator confluence scoring and 30-minute TTL
- Backtester — natural-language strategy description → LLM-parsed entry/exit rules → NumPy simulation → equity curve + Sharpe/drawdown/win-rate metrics
- AI Command Palette (
⌘K) — type any trade in plain English ("buy $2k of NVDA and trim TSLA by 20%") → structured step-by-step plan → one-click execution - Correlation Network — physics-based force-directed graph of portfolio holdings colored by sector and weighted by correlation strength
- 3-layer cache — in-memory (60s) → Redis → Finnhub REST; respects free-tier rate limits (60 req/min)
- Async SQLAlchemy 2.0 — all DB operations non-blocking;
with_for_update()on trades for atomic P&L - JWT auth — register/login → Bearer token →
Authorizationheader injected byapiFetch
- Frontend: Next.js 15, shadcn/ui, Magic UI, Framer Motion, TradingView Charts
- Backend: FastAPI, SQLAlchemy 2.0 (async), PostgreSQL, Redis
- AI/ML: XGBoost + LSTM + LogReg ensemble, FinBERT + Gemini sentiment, LangChain agent, ChromaDB RAG
- Real-time: Finnhub WebSocket, SSE streaming