💻 computer science

Sentinelbi AI: an AI-Powered Business Intelligence System for Revenue Monitoring and Anomaly Detection

This paper introduces SentinelBI AI, a deployable business intelligence platform that integrates Isolation Forest and LSTM Autoencoder models to detect anomalous e-commerce revenue patterns, fusing their outputs into a three-tier alert system accessible via an interactive Streamlit dashboard for actionable decision support.

Original authors: Tata Venkata Karthik, B. Srinivas S. P. Kumar

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tata Venkata Karthik, B. Srinivas S. P. Kumar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the bustling world of online shopping, a steady stream of transactions flows every second, creating a massive, complex record of what people buy, when they buy it, and how much they spend. For business owners, this data is a treasure trove, but it is also a challenge. Most small and mid-sized companies rely on simple charts that show total sales, much like a speedometer in a car that only displays the current speed. These charts tell a manager how fast they are going right now, but they do not warn them if the engine is sputtering or if the car is suddenly veering off course. The real danger lies in the unexpected: a sudden, unexplained spike in sales that might indicate a pricing error, a quiet dip that suggests a stock shortage, or a strange pattern that hints at fraud. To catch these issues, businesses need more than just a snapshot of the present; they need a system that understands what "normal" looks like over time and can instantly spot when the rhythm of business breaks. This is the domain of anomaly detection, a field where computers learn to recognize the usual patterns of data so they can raise an alarm the moment something unusual happens.

Two researchers from Chaitanya Bharathi Institute of Technology in India have built a new tool designed to solve this exact problem for online retailers. They call their creation SentinelBI AI, a system that acts as a vigilant guardian for revenue streams. Instead of relying on a single, rigid rule to decide what is wrong, the system uses two different types of artificial intelligence working together. One type is good at spotting strange individual days, like a sudden, isolated spike in sales that doesn't fit the day's typical profile. The other type is better at understanding the flow of time, recognizing when a sequence of days follows a weird pattern that a human might miss. By combining the insights from both, the system creates a three-level warning scale: normal, warning, and critical. This approach allows the system to filter out the noise of everyday fluctuations and focus only on the days that truly need attention.

The researchers tested their system using a real-world dataset from a major online gift retailer, covering over a million individual transactions from late 2009 to late 2011. They cleaned this massive log of data and organized it into 731 daily summaries, tracking key metrics like total revenue, the number of items sold, the count of transactions, and the number of active customers. They also calculated how much each day's revenue differed from the average of the previous week. With this prepared information, they trained their two artificial intelligence models. The first model, known as an Isolation Forest, learned to identify unusual days by looking at the specific combination of numbers for that single day. The second model, an LSTM Autoencoder, learned the normal shape of revenue over a thirty-day window, allowing it to detect when the recent history of sales looked wrong, even if the numbers on a single day seemed okay.

When the system ran its analysis on the 731 days of business history, it found that the two models agreed on the vast majority of days, correctly identifying most of the time as normal business. However, the real value appeared in the days where the models disagreed or where both sounded the alarm. The system flagged 19 days as "Critical," meaning both models agreed something was seriously wrong. It identified another 32 days as "Warning," where one model sounded the alarm but the other did not. The remaining 680 days were classified as normal. This dual-model approach proved to be more accurate than a simple, static rule that just looks for numbers that are too high or too low. While a basic rule would have flagged 41 days with a lower success rate of matching actual unusual events, the researchers' smarter system flagged fewer days but caught a higher percentage of the real problems, with the two models achieving match rates of over 50 percent.

To make these findings useful for actual business people, the researchers did not stop at the computer code. They built a complete, interactive dashboard that anyone can use in a web browser, without needing to know how to program. This dashboard has five different pages. One page gives a high-level overview for executives, while others dive deep into the specific anomalies. A particularly innovative feature is a natural-language query tool. This allows a user to type a question in plain English, such as asking for the top-selling products in a specific country, and the system automatically translates that question into the necessary database commands to find the answer. This removes the barrier of technical jargon, letting managers ask questions and get answers instantly.

The study revealed some clear insights about how to best monitor a business. The most important factor in spotting an anomaly was not the total amount of money made, but how much that day's revenue differed from the recent past. A day with high sales is not necessarily a problem if it fits the pattern of the week before; a day is only strange if it breaks the rhythm. The system also showed that the United Kingdom, which was the primary market for the data, dominated the revenue, suggesting that anomalies in high-volume countries deserve the closest attention. By fusing the strengths of two different artificial intelligence methods, the researchers created a tool that is both precise and practical. It turns complex, unsupervised machine learning into a simple, actionable guide, helping businesses move from simply watching their sales numbers to actively understanding and protecting their financial health.

Technical Summary: SentinelBI AI

Problem Statement
Modern e-commerce operators generate continuous streams of transactional data, yet many small and mid-sized platforms rely on static dashboards that report totals without actively flagging deviations from normal business behavior. Sudden revenue spikes, prolonged dips, or unusual combinations of transaction metrics can indicate critical issues such as pricing errors, fraud, stock-outs, or data pipeline faults. However, these anomalies often go unnoticed until it is too late. The paper identifies a gap in existing systems: the lack of an integrated solution that combines business intelligence reporting with unsupervised anomaly detection to provide timely, actionable alerts for revenue monitoring.

Methodology
The authors propose SentinelBI AI, an end-to-end business intelligence platform designed to detect anomalous revenue behavior using a dual-model unsupervised approach. The system is built upon the Online Retail II dataset (UCI), comprising 1,067,371 transaction records from December 2009 to December 2011 across 43 countries.

The technical pipeline proceeds as follows:

Data Preprocessing & Feature Engineering: Raw transaction logs are cleaned (removing cancellations and negative quantities) and aggregated into 731 daily business-performance records. Five key features are engineered: Total Revenue, Total Quantity, Transaction Count, Active Customer Count, and Revenue Deviation (the difference between actual daily revenue and its 7-day rolling average).
Dual-Model Detection: Two complementary unsupervised detectors are trained on the standardized daily features:
- Isolation Forest: An ensemble method that isolates anomalies by randomly partitioning the feature space. It is tuned with a 5% contamination rate to flag unusual individual daily observations (point anomalies).
- LSTM Autoencoder: A deep learning model trained on 30-day sequential windows of revenue data. It learns the normal temporal shape of revenue and flags days where the reconstruction error exceeds a threshold ( $\mu + 2\sigma$ ), effectively detecting contextual or collective anomalies in the time series.
Alert Fusion: The outputs of both models are fused into a three-tier alert scheme:
- Critical: Flagged by both models.
- Warning: Flagged by exactly one model.
- Normal: Flagged by neither.
Deployment: The pipeline is deployed via a Streamlit dashboard featuring five pages: Executive Overview, Business Analytics, AI Anomaly Monitor, AI Investigation, and a natural-language query interface ("Ask SentinelBI") backed by a SQLite analytical warehouse.

Key Results
On the dataset of 731 business days, the system produced the following results:

Detection Rates: The Isolation Forest independently flagged 37 days (5.06%), and the LSTM Autoencoder flagged 33 days (4.51%).
Fusion Performance: 19 days were jointly flagged and classified as Critical (2.60% of total days), while 32 days were classified as Warning (4.38%). The remaining 680 days were Normal.
Model Comparison: When compared against a static $\pm 2\sigma$ $\pm 2 σ$ baseline, the learned detectors demonstrated superior performance in cross-validated match rates against manually reviewed unusual days:
- Isolation Forest: 51.4% match rate.
- LSTM Autoencoder: 57.6% match rate.
- Static Baseline: 43.9% match rate.
Feature Importance: A sensitivity analysis revealed that Revenue Deviation was the most significant feature (contributing 34% to the Isolation Forest's discriminative power), confirming that deviation from a rolling baseline is a stronger anomaly signal than raw absolute scale.

Significance and Contributions
The paper claims its primary contribution is the development of a deployable, dual-model anomaly detection pipeline that successfully converts unsupervised model outputs into an actionable, non-technical decision-support tool.

Key aspects of this significance include:

Complementary Detection: By combining a point-anomaly detector (Isolation Forest) with a sequence-anomaly detector (LSTM), the system reduces false positives and captures different types of anomalies that a single model might miss.
Operational Usability: The integration of the detection pipeline into a browser-based Streamlit dashboard with a natural-language query layer allows operations staff and non-technical stakeholders to investigate revenue anomalies without needing to write code or SQL.
Actionable Prioritization: The three-tier alert system (Normal, Warning, Critical) helps operations teams prioritize investigative efforts, focusing resources on the small subset of days (2.6%) most likely to be genuinely unusual.

The authors conclude that while the system is currently unsupervised, it provides a robust foundation for revenue monitoring that outperforms static statistical rules, offering a practical solution for timely anomaly detection in e-commerce environments.

Technical Summary: SentinelBI AI

More like this