HackIllinois 2026 — Kalshi News Trading Bot

An autonomous trading system that reads real-time financial news, runs multi-stage NLP/LLM analysis on GPU-accelerated cloud infrastructure, and executes limit orders on Kalshi prediction markets.

---

How It Works

RSS Feeds (25+)
      │
      ▼
  News Ingestion              News/rss.py
      │  deduplicated by URL hash
      ▼
  Ticker Matching             NLP/ticker_modal.py
      │  SentenceTransformers (all-MiniLM-L6-v2) on Modal T4 GPU
      │  matches headlines → Kalshi market tickers
      ▼
  Financial Sentiment         NLP/sentiment.py
      │  ProsusAI/FinBERT on Modal A10G GPU
      │  returns label + confidence score
      ▼
  LLM Signal Routing          LLM/llm_signal.py
      │  Groq API · Llama 3.3 70B
      │  resolves ambiguous signals with full article context
      ▼
  Order Execution             Kalshi/kalshi_order_executor.py
      │  RSA-PSS signed REST requests
      │  places YES/NO limit orders at best ask
      ▼
  Sell Heartbeat              Kalshi/sell_heartbeat.py
         background thread · monitors open positions · closes on reversal

---

Tech Stack

Component	Technology
Financial Sentiment	ProsusAI/FinBERT
Market Matching	SentenceTransformers all-MiniLM-L6-v2
GPU Inference	Modal — A10G (FinBERT) · T4 (embeddings)
LLM Routing	Groq API — Llama 3.3 70B Versatile
Prediction Markets	Kalshi REST API — RSA-PSS authenticated
News Ingestion	feedparser — 25+ RSS feeds
Backend	Flask + SSE (Server-Sent Events)
Frontend	React + Vite + Three.js
Frontend Hosting	Vercel
Backend Hosting	Railway

---

Project Structure

.
├── main.py                       # Orchestration loop
├── api/
│   └── index.py                  # Flask API (start/pause/status/logs/news SSE)
├── News/
│   └── rss.py                    # RSS polling + deduplication
├── NLP/
│   ├── ticker_modal.py           # Modal-powered market ticker matching
│   └── sentiment.py              # Modal-powered FinBERT sentiment
├── LLM/
│   └── llm_signal.py             # Groq/Llama 3.3 70B signal resolver
├── Kalshi/
│   ├── kalshi_auth.py            # RSA-PSS request signing
│   ├── kalshi_order_executor.py  # Limit order placement
│   ├── market_utils.py           # Best ask price lookup
│   └── sell_heartbeat.py         # Background position monitor
├── Frontend/
│   └── HackIllinois-2026/        # React + Vite app
├── sentiment_output.csv          # Pipeline output log
├── requirements.txt
├── Procfile                      # Railway entrypoint
└── vercel.json                   # Vercel frontend build config

---

Local Development

Prerequisites

• Python 3.11+
• Node.js 18+
• A Modal account (modal token new)
• A Groq API key
• Kalshi API key + RSA private key

Backend

# Install dependencies
pip install -r requirements.txt

# Create a .env file at the repo root
cat > .env <<EOF
GROQ_API_KEY=your_groq_key
KALSHI_API_KEY=your_kalshi_key
KALSHI_PRIVATE_KEY="-----BEGIN PRIVATE KEY-----\n...\n-----END PRIVATE KEY-----"
EOF

# Start the Flask API server (port 8000)
python api/index.py

Frontend

cd Frontend/HackIllinois-2026
npm install
npm run dev          # starts Vite dev server on port 5173

Open http://localhost:5173. The Vite proxy forwards /api/* to localhost:8000.

Click Start in the dashboard to launch the trading pipeline.

---

Deployment

Backend → Railway

1. Go to railway.app → New Project → Deploy from GitHub repo
2. Set environment variables in the Railway dashboard:
   • GROQ_API_KEY
   • KALSHI_API_KEY
   • KALSHI_PRIVATE_KEY
   • PORT=8000
3. Railway reads Procfile and runs python api/index.py automatically.
4. Copy the Railway public URL (e.g. https://your-app.railway.app).

Frontend → Vercel

1. Go to vercel.com → New Project → import this GitHub repo
2. Set the following in Vercel project settings:
   • Build command: cd Frontend/HackIllinois-2026 && npm install && npm run build
   • Output directory: Frontend/HackIllinois-2026/dist
   • Environment variable: VITE_API_BASE=https://your-app.railway.app
3. Deploy. Vercel serves the static frontend; all API calls route to Railway.

Note: The Flask backend cannot run on Vercel — it requires long-lived processes, subprocess spawning, file writes, and packages (torch, transformers) that exceed Vercel's 250 MB bundle limit. Modal GPU inference runs on Modal's own cloud regardless of where the backend is hosted.

---

Environment Variables

Variable	Where	Description
GROQ_API_KEY	Backend	Groq API key for Llama 3.3 70B
KALSHI_API_KEY	Backend	Kalshi exchange API key
KALSHI_PRIVATE_KEY	Backend	PEM-encoded RSA private key for request signing
PORT	Backend	Flask listen port (default 8000)
VITE_API_BASE	Frontend (build)	Railway backend URL (empty string for local dev)

---

Pipeline Configuration

Edit constants at the top of main.py:

POLL_INTERVAL_S       = 10    # seconds between RSS polls
MIN_FINBERT_SCORE     = 0.70  # minimum FinBERT confidence to act on
MIN_TICKER_CONFIDENCE = 0.40  # minimum market match confidence to act on

Order sizing and take-profit/stop-loss thresholds are configurable via the dashboard UI at runtime (sent to /api/thresholds).