Multi-strategy AI trading agent demonstrating a modular architecture for quantitative research, AI-assisted analysis, and automated portfolio monitoring.
This Public Edition (SQ-PE) is a sanitized showcase version intended for engineering demonstration and portfolio purposes. Certain components such as proprietary signals, calibrated thresholds, production broker integrations, and optimized portfolio parameters are intentionally omitted or simplified.
The goal of this repository is to demonstrate systems architecture, AI integration, and quantitative engineering practices without exposing proprietary trading logic.
This project is accompanied by the paper:
Sentinel Quant: A Human-in-the-Loop Multi-Model Portfolio Management System with Formal AI Governance Architecture https://doi.org/10.5281/zenodo.19284847
The public dashboard provides real-time visibility into the system's performance, including:
- Portfolio equity curve with S&P 500 (SPY) benchmark comparison
- Alpha vs SPY calculation showing risk-adjusted outperformance
- Position management across three strategy sleeves
- Risk governance state (regime, VIX, exposure, beta)
- AI intelligence calibration and accuracy metrics
- Quarterly ETF review with recommendation tracking
- Trade journal with AI post-mortem analysis
- Fund oversight and intelligence review cycle
+------------------+
| Market Hours |
| Gate |
+--------+---------+
|
+--------v---------+
| Market Screener |
| Periodic scans |
+--------+---------+
|
+--------------+--------------+
| | |
+--------v---+ +------v-----+ +-----v------+
| Short-Term | | Mid-Term | | Long-Term |
| Momentum | | Trend | | Portfolio |
| Signals | | Signals | | Allocation |
+--------+---+ +------+-----+ +-----+------+
| | |
+--------------+--------------+
|
+--------v---------+
| Regime Detection |
| Market Volatility|
+--------+---------+
|
+--------------+--------------+
| | |
+--------v---+ +------v-----+ +-----v------+
| Risk | | Volatility | | Beta |
| Engine | | Management | | Monitoring |
+--------+---+ +------+-----+ +-----+------+
| | |
+--------------+--------------+
|
+--------v---------+
| Intelligence |
| Calibration + |
| Meta-Model |
+--------+---------+
|
+--------------+--------------+
| | |
+--------v---+ +------v-----+ +-----v------+
| Execution | | Dashboard | | Notification|
| Sim Broker | | Next.js | | System |
+------------+ +------------+ +-------------+
The system manages three concurrent strategy sleeves with independent risk budgets: short-term momentum, mid-term trend following, and long-term strategic allocation via quarterly ETF review. Each sleeve operates with its own entry criteria, holding period, and exit logic.
Every trade proposal passes through a multi-layered risk engine before execution. The system uses a human-in-the-loop governance model where AI generates recommendations and a formal risk framework enforces constraints. No trade executes without passing all governance checks.
The risk governance system implements a multi-layered protection framework with 16+ automated checks covering:
- Order integrity — duplicate prevention, daily trade limits
- Loss containment — global and per-sleeve daily loss caps
- Position management — sleeve-level count and size limits, minimum trade thresholds
- Earnings protection — calendar-aware blocking around earnings announcements with configurable pre- and post-earnings windows
- Portfolio protection — drawdown circuit breaker, global exposure cap, cash reserve enforcement
- Diversification — sector concentration limits at both portfolio and sleeve level, correlation-based blocking for overlapping positions
- Entry quality — technical confirmation scoring, intraday momentum validation
- Adaptive sizing — loss streak detection with automatic position size reduction, regime-aware confidence adjustment
- Sector intelligence — momentum-based confidence adjustments favouring strong sectors and penalising weak ones
- AI calibration — confidence threshold enforcement based on model self-assessment
Short-term positions are managed with automated exit rules including stop-loss, take-profit, and time-based exits at calibrated thresholds.
Blocks buy proposals near scheduled earnings announcements with configurable time windows per strategy sleeve. Includes a post-earnings cool-off period to prevent chasing initial reactions. Uses free market calendar data with intelligent caching to minimise API calls.
Monitors performance across all major market sectors using ETF proxies. Sectors showing strong relative momentum receive a confidence boost on buy proposals, while underperforming sectors receive a confidence penalty. Data is cached and injected into strategy prompts for both short-term and mid-term analysis.
When the AI recommends buying a stock, the original reasoning is stored alongside the position. When evaluating whether to sell, the original buy thesis is injected into the prompt so the AI can assess whether the thesis has broken or still holds, rather than reacting purely to price movement.
After every closed trade, the AI analyses the decision with structured scoring across multiple dimensions. Results are stored in an append-only journal that builds a dataset for identifying systematic biases and improving future decision-making.
Recent trade outcomes are injected into strategy prompts, allowing the AI to calibrate its confidence based on its own track record. The system tracks performance by sleeve and adjusts behaviour based on accumulated evidence.
Before approving a buy, the system evaluates multiple technical confirmations. Trades that fail to meet the quality threshold receive a confidence penalty, reducing position size on marginal setups rather than hard-blocking them.
If the portfolio experiences a significant intraday drawdown, all new buy proposals are temporarily suspended. This prevents the system from buying into a falling market and compounding losses during adverse conditions.
After a configurable number of consecutive losing trades, position sizes are automatically reduced until the next winning trade resets the counter. This proven risk management technique prevents compounding during cold streaks.
The equity curve displays a dual-axis comparison: portfolio value vs S&P 500 percentage return with a toggle button. Returns are deposit-adjusted so periodic capital additions don't inflate apparent performance vs the benchmark.
Displays portfolio return minus SPY return as a single metric, showing risk-adjusted outperformance at a glance.
When post-mortem entries exist, the dashboard displays scored journal entries with aggregate statistics and individual trade analyses, providing transparency into the AI's self-assessment process.
The system generates multiple automated reports on configurable schedules:
| Report | Description |
|---|---|
| Weekly Trading Report | Trade summary, P&L breakdown, risk events |
| API Usage Report | AI API call breakdown, cost analysis |
| Weekend Portfolio Digest | Opus-powered narrative analysis with separate private and public editions |
| Monthly Review | Long-term trend analysis, strategy effectiveness assessment |
| Daily Lesson | Post-market single-paragraph lesson from the day's trading activity |
The system runs on cloud infrastructure with containerised trading agents, a modern web dashboard, and scheduled intelligence tasks.
| Component | Technology |
|---|---|
| Trading agents | Python, Docker |
| Dashboard | Next.js, Recharts, Tailwind CSS |
| AI models | Claude Sonnet (trading), Claude Opus (reviews) |
| Hosting | Cloud VPS (agents), Vercel (dashboard) |
| Data | Yahoo Finance, Alpaca Markets API |
| Feature | Public Edition | Production System |
|---|---|---|
| Broker | Simulated environment | Live broker integrations |
| Signals | Simplified / partial | Proprietary signal models |
| Risk parameters | Generic examples | Calibrated portfolio parameters |
| Universe size | Reduced | Full research universe |
| AI prompts | Simplified | Tuned and optimized |
| Infrastructure | Demonstration | Production deployment |
- Zero-trust architecture for AI-generated signals
- Multi-layered risk engine with 16+ automated governance checks
- Earnings-aware trading with calendar integration
- Sector momentum rotation with dynamic confidence adjustment
- AI self-reflection via trade journal post-mortems
- Deposit-adjusted benchmark comparison against S&P 500
- Regime-adaptive portfolio behaviour
- AI confidence calibration and self-correction
- Automated drawdown protection and loss streak management
- Modular quant research framework
The system exposes a public metrics endpoint serving anonymized, real-time performance data:
GET https://sentinel-quant-dashboard.vercel.app/api/metrics
Response (JSON, CORS-enabled):
| Field | Description |
|---|---|
| performance.total_return_pct | Cumulative portfolio return |
| performance.sharpe_ratio | Risk-adjusted return metric |
| performance.win_rate_pct | Percentage of profitable trades |
| performance.total_trades | Lifetime trade count |
| risk.regime | Current market regime (LOW_VOL / NORMAL / HIGH_VOL / CRISIS) |
| risk.vix_level | Current VIX reading |
| risk.portfolio_beta | Portfolio beta vs benchmark |
No authentication required. Position details and trade reasoning are stripped for privacy. Data refreshes every 30 minutes during market hours.
Trading financial markets involves substantial risk. This repository is intended for research and educational purposes only. The public implementation is not intended for live trading and does not represent financial advice.
If you reference this project or its architecture, please cite:
Oke, I. E. (2026). Sentinel Quant: A Human-in-the-Loop Multi-Model Portfolio Management System with Formal AI Governance Architecture (Version 1.0). Zenodo. https://doi.org/10.5281/zenodo.19284847
@report{oke2026sentinelquant,
author = {Oke, Iyanuoluwa Enoch},
title = {Sentinel Quant: A Human-in-the-Loop Multi-Model Portfolio Management System with Formal AI Governance Architecture},
year = {2026},
publisher = {Zenodo},
version = {1.0},
doi = {10.5281/zenodo.19284847},
url = {https://doi.org/10.5281/zenodo.19284847}
}MIT License
Sentinel Quant - Public Engineering Showcase Developed by Iyanuoluwa Oke
GitHub: https://github.com/Iyanuoluwa007
If you explore the project and have suggestions or comments, feedback is welcome:
Google Form: https://forms.gle/jKDN18pMzP3mCpJU6
Feedback from researchers, engineers, and quantitative developers is especially appreciated.
Released under the MIT Licence. See LICENSE.
Copyright (c) 2026 Oke Iyanuoluwa Enoch.
You are free to use, modify and build on this work. The licence asks one thing in return, and it is not optional: keep the copyright notice and the licence text with any copy or substantial portion of the software. That notice is how the work stays credited to its author.
If you fork this, publish something derived from it, or use it in a product, paper or demo, please credit it visibly and link back:
Built on Sentinel-Quant_PE by Oke Iyanuoluwa Enoch.
The MIT Licence covers this repository own code only. Any third-party data, pretrained weights, papers or assets that may be present remain the property of their owners and are subject to their own terms.