QuantumSentinel

💡 Inspiration

Every year, billions of dollars are lost to stock market manipulation — pump and dump schemes, short squeezes, and anomalous trading activity that classical algorithms consistently miss. Traditional z-score and moving average detectors look at price and volume in isolation. We asked a different question: what if quantum interference patterns could detect anomalies that classical math simply can't see?

The 2021 GameStop short squeeze was our north star. A Reddit community took on Wall Street hedge funds and won — sending GME from $20 to $483 in two weeks. Every classical detector missed the early warning signs. We wanted to build something that wouldn't.

⚛️ What We Built

QuantumSentinel is a Gemini-powered AI agent that detects anomalous stock market activity using a Variational Quantum Classifier (VQC) built with PennyLane. Every decision the agent makes is traced in Arize Phoenix — giving the system full observability and a self-improvement loop.

How it works

1. Data Layer — Fetches live OHLCV data from Yahoo Finance and engineers 13 technical features including RSI, Bollinger Band position, MACD histogram, and volume ratios

2. Quantum Layer — A 4-qubit VQC encodes features as quantum rotation angles:

$$|\psi\rangle = R_Y(\pi x_i)|0\rangle$$

Variational layers with entanglement learn anomaly boundaries in quantum feature space. The circuit outputs an expectation value:

$$\langle Z_0 \rangle = \langle \psi | Z \otimes I \otimes I \otimes I | \psi \rangle$$

Scores below the learned threshold are classified as anomalies.

1. Agent Layer — A Gemini agent orchestrates three tools:

   • fetch_stock_data — pulls live technical indicators
   • quantum_scan — runs the VQC and returns anomaly score
   • get_trace_summary — queries Arize Phoenix for self-improvement

2. Observability Layer — Every quantum inference and Gemini call is traced in Arize Phoenix via OpenInference instrumentation. The agent periodically reads its own trace history and adjusts its detection threshold automatically.

🚀 The GME Proof Point

We backtested QuantumSentinel on the January 2021 GameStop short squeeze — one of the most extreme anomalous trading events in modern history.

The quantum circuit caught every single squeeze day. Classical missed half.

This isn't coincidence — the VQC learns decision boundaries in quantum feature space that classical linear separators cannot express. The interference patterns between entangled qubits capture correlations across all 13 features simultaneously.

🔧 How We Built It

• Quantum ML — PennyLane VQC with 4 qubits, 2 variational layers, Adam optimizer, trained per-ticker on 1 year of historical data
• AI Agent — Google Gemini 2.5 Flash via google-genai SDK with function calling
• Observability — Arize Phoenix + OpenInference instrumentation, custom quantum inference spans
• Backend — FastAPI + Uvicorn
• Deployment — Google Cloud Run
• Data — Yahoo Finance API (yfinance)

😤 Challenges

The hardest challenge was making quantum meaningful. It's easy to bolt a quantum circuit onto a project as decoration. Making it genuinely contribute to better anomaly detection — and proving it with the GME benchmark — required careful feature engineering and threshold calibration.

Per-ticker training was another challenge. A model trained on AAPL data would flag TSLA as anomalous simply because it's different — not because it's suspicious. We built a per-ticker weight system that trains each stock on its own historical patterns.

Arize Phoenix on Windows required running Phoenix as a separate process rather than embedded, which led to a complete rewrite of our tracing architecture.

📚 What We Learned

• Variational Quantum Classifiers are genuinely viable for financial anomaly detection — not just academic curiosities
• AI observability isn't optional — being able to see every agent decision in Phoenix was crucial for debugging and improving the system
• The self-improvement loop (agent querying its own traces) is the most underexplored capability in modern AI agents

🔮 What's Next

• Deploy on real quantum hardware (IBM Quantum, IonQ) for true quantum speedup
• Extend to crypto markets and options flow
• Real-time alerts via email/SMS when anomalies are detected
• Multi-stock correlation analysis using quantum entanglement

Built With

• arize-phoenix
• fastapi
• google-adk
• google-cloud-run
• google-gemini
• numpy
• openinference
• opentelemetry
• pennylane
• python
• scikit-learn
• yahoo-finance-api