Issuu

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

SIM-OPS: An Autonomous Multi-Agent Framework for Intelligent Business Operations Management Arshvir Singh Kalsi1, Viraj Prabhu2, Ghrani Poojari3, Hasti Shah4, Varsha Nagpurkar5 1,2,3,4 Department of Computer Engineering, St. Francis Institute of Technology, Mumbai, India 5Professor, Department of Computer Engineering, St. Francis Institute of Technology, Mumbai, India

---------------------------------------------------------------------***---------------------------------------------------------------------

Abstract-This

components to execute complex workflows without human intervention.

paper presents SIM-OPS (Smart Intelligence Management for Operations Planning & Supervision), an autonomous multi-agent system designed to function as an intelligent business operations manager. The system leverages advanced machine learning models (XGBoost, Isolation Forests), real-time data processing, and coordinated agent orchestration powered by Google Gemini via the LangChain framework to continuously monitor organizational metrics, predict business risks, and execute automated interventions. The framework integrates six specialized agents – Monitoring, Prediction, Decision, Action, Reporting and Feedback – that communicate through a database-backed message bus to perform end-toend business intelligence operations. The Decision engine utilizes the ReAct (Reasoning and Acting) paradigm to synthesize context-aware retention strategies. The system achieves 86.94% accuracy in churn prediction with 0.89 average confidence, processes predictions autonomously every 6 hours, and executes multi-channel alerts through Slack, Jira and email. Experimental results demonstrate that the autonomous agent chain reduces manual intervention by 90% while achieving a 96% autonomous decision rate and a 0.3 second average response time for critical alerts. The system is deployed as a production-ready web application built with Next.js 16, TypeScript, Supabase and LangChain, with a Python-based FastAPI machine learning service.

1.1 Motivation Delayed responses to churn signals, revenue anomalies and operational inefficiencies directly translate into lost revenue and wasted analyst hours. SIM-OPS is motivated by the need for a system that can observe business metrics around the clock, reason about which events matter, and execute the appropriate action without waiting on a human operator to be available.

1.2 Contributions This paper makes the following contributions: (1) a novel six-agent framework in which specialized agents collaborate through a message bus to perform comprehensive business operations management; (2) a fully automated prediction pipeline that runs machine learning inference every 6 hours and triggers agent chains based on risk thresholds; (3) an intelligent, ReAct-based decision engine powered by Google Gemini that evaluates predictions, applies business logic and determines appropriate actions with 91% average confidence; (4) a multi-channel action executor that delivers alerts through Slack, creates Jira tickets and sends emails based on severity classification; and (5) a production-ready implementation deployed with modern web technologies, achieving sub-second response times and 24/7 autonomous operation.

Key Words: Multi-agent systems, autonomous operations, machine learning, business intelligence, churn prediction, workflow automation, real-time monitoring

2. LITERATURE REVIEW

1. INTRODUCTION

2.1 Multi-Agent Systems

Modern businesses generate vast amounts of operational data from customer interactions, financial transactions, support systems and product usage. Extracting actionable insights from this data and executing timely interventions remains a significant challenge. Traditional business intelligence systems require constant human oversight, leading to delayed responses to critical events such as customer churn, revenue anomalies and operational inefficiencies. The need for autonomous, intelligent systems that can continuously observe, analyze and act on business data has become increasingly critical. Such systems must not only detect patterns and predict outcomes but also coordinate multiple specialized

Impact Factor value: 8.315

Multi-agent systems (MAS) have been extensively studied in artificial intelligence research [1], [14]. Traditional MAS frameworks, as surveyed by Ferber [15] and Stone and Veloso [16], focus on agent communication protocols, coordination mechanisms and distributed problem-solving. Multi-agent reinforcement learning [17] has advanced cooperative and competitive agent behaviors; however, most existing systems operate in simulated environments or require significant human oversight. Recent work has explored reinforcementlearning approaches [2] and hierarchical agent

ISO 9001:2008 Certified Journal

Page 72

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

3.1 Predictive Modeling with XGBoost

architectures [3]. SIM-OPS differs by focusing on practical business operations with real-world constraints, emphasizing reliability, explainability and integration with existing enterprise systems.

The core churn-prediction model utilizes XGBoost (eXtreme Gradient Boosting), an optimized distributed gradient-boosting library. XGBoost minimizes a regularized objective function:

2.2 LLM-Based Autonomous Agents The emergence of Large Language Models (LLMs) has catalyzed a new generation of autonomous agent architectures. Park et al. [18] demonstrated that LLMpowered generative agents can exhibit believable, emergent social behaviors through memory and planning modules. Xi et al. [19] provide a comprehensive survey of LLM-based agents, categorizing them by perception, reasoning and action capabilities. SIM-OPS builds on these paradigms by deploying LLM-based agents in a production business-operations context, where reliability, auditability and integration with external enterprise systems (Slack, Jira, email) are primary constraints rather than openended generative behavior.

L(t) = Σᵢ l(yᵢ, ŷᵢ(t−1) + fₜ(xᵢ)) + Ω(fₜ)

…(1)

where l is a differentiable convex loss function measuring the difference between the prediction ŷᵢ and the target yᵢ. The term Ω penalizes the complexity of the model:

Ω(f) = γT + ½λ‖w‖²

…(2)

with T the number of leaves in the tree and w the leaf weights. This regularization helps prevent overfitting, a common issue in customer behavioral datasets with high dimensionality.

2.3 Business Intelligence and Predictive Analytics

3.2 Anomaly Detection via Isolation Forests

Predictive analytics in business intelligence has traditionally relied on batch processing and periodic reporting [4]. Modern approaches incorporate real-time streaming analytics [5] and automated machine learning (AutoML) [6]. Data-mining techniques have been widely applied to customer relationship management [20]. Customer churn prediction has been addressed using logistic regression [7], random forests [8], [21] and gradient boosting [9]; comparative studies [22] show that ensemble methods consistently outperform single classifiers in churn-prediction tasks. SIM-OPS extends these approaches by integrating predictions into an autonomous action-execution framework.

For monitoring KPIs such as revenue drops or sudden spikes in support tickets, Isolation Forests [23] are employed. Unlike traditional methods that profile normal points, Isolation Forests explicitly isolate anomalies. The path length h(x) to isolate a point x is the number of edges from the root node to a terminating node, and the anomaly score is defined as:

s(x, n) = 2 −E(h(x)) / c(n)

where E(h(x)) is the average path length and c(n) is the average path length of unsuccessful searches in a Binary Search Tree. Anomalies are identified when s(x, n) approaches 1, indicating that the point is easily isolated and thus abnormal.

2.4 Workflow Automation Workflow automation systems such as Apache Airflow [10] and Temporal [11] provide orchestration capabilities but require manual workflow definition and lack intelligent decision-making. Business process management (BPM) systems [12] offer structured workflow execution but do not incorporate machine learning or autonomous agent coordination. SIM-OPS combines workflow automation with intelligent agents that dynamically determine appropriate actions based on real-time predictions and business rules.

3.3 LLM-Based Agentic Reasoning The multi-agent orchestration relies on Large Language Models employing the ReAct (Reasoning and Acting) framework [13]. ReAct prompts the LLM to generate both reasoning traces and task-specific actions in an interleaved manner. Given a context c, the agent generates a sequence of thoughts tᵢ, actions aᵢ and observations oᵢ:

τᵢ = LLM(c, t₁, a₁, o₁, …, tᵢ₋₁, aᵢ₋₁, oᵢ₋₁)

3. THEORETICAL BACKGROUND

Impact Factor value: 8.315

…(4)

This paradigm allows the Decision Agent to dynamically evaluate the context of a high-risk customer and synthesize a tailored retention strategy rather than relying solely on static rule-based systems.

To provide a robust foundation for SIM-OPS, principles from predictive machine learning, anomaly detection and modern natural language processing are synthesized below.

…(3)

ISO 9001:2008 Certified Journal

Page 73

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

4. SYSTEM ARCHITECTURE

30-day revenue forecasts with confidence intervals. All models are pre-trained and stored as pickle files, loaded on demand for inference, and exposed through RESTful API endpoints. Fig -2 shows the predictive-analytics interface.

4.1 Overview SIM-OPS employs a layered architecture consisting of four primary components: (1) Data Ingestion Layer, (2) Machine Learning Service, (3) Multi-Agent Orchestration Layer, and (4) Action Execution Layer, shown in Fig -1.

Fig -2: Predictive analytics interface displaying ML predictions and churn risk

4.4 Multi-Agent Orchestration Layer

Fig -1: SIM-OPS system architecture

The core innovation of SIM-OPS is its multi-agent orchestration system consisting of six specialized agents. The live operational status, active workflows and agent execution statistics are monitored via the unified dashboard shown in Fig -3.

4.2 Data Ingestion Layer The data ingestion layer collects data from multiple sources: Stripe webhooks carrying real-time payment events (successful transactions and payment failures); the customer database holding user profiles, subscription data and behavioral metrics in PostgreSQL (Supabase); analytics events covering product usage, feature adoption and engagement metrics; and support systems providing ticket counts, resolution times and customer-satisfaction scores. Webhook handlers process incoming events and update the customer database in real time, ensuring that the machine learning models always operate on current data.

4.3 Machine Learning Service The ML service is implemented as a Python FastAPI application providing four primary models. The Churn Prediction Model uses an XGBoost classifier trained on tenure, support-ticket count, payment-failure history, feature-usage rate, session frequency and duration, and total spend and discount usage; it achieves 86.94% accuracy with 0.89 average confidence on the held-out test set (80/20 stratified split). The Customer Lifetime Value (CLV) Model predicts expected revenue per customer using gradient-boosting regression on purchase history, average order value, purchase frequency, engagement score and referral activity. The Anomaly Detection Model implements the Isolation Forest algorithm to detect unusual patterns in time-series metrics such as revenue, user growth and system costs. The Revenue Forecasting Model uses the Prophet time-series model [24] to generate

Impact Factor value: 8.315

Fig -3: Autonomous multi-agent monitoring and configuration portal The Monitoring Agent is responsible for KPI-deviation detection and threshold monitoring: it continuously monitors business metrics (churn rate, revenue, user growth), compares current values against configured thresholds (e.g. churn rate 3.0%, revenue deviation 5.0%, user growth −2.0%), detects anomalies using statistical methods, and triggers alerts when thresholds are exceeded. The Prediction Agent performs ML inference and risk scoring: it fetches customer data, extracts features for the ML models, executes churn and CLV predictions, stores

ISO 9001:2008 Certified Journal

Page 74

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

4.6 Action Execution Layer

predictions with confidence scores, and identifies highrisk customers (churn probability greater than 80%). It processes 100 customers in 2.1 seconds on average.

The action execution layer implements integrations with external systems. The Slack integration uses a webhook API to send formatted alerts with colour-coded severity indicators, customer details, risk scores and recommended actions. The Jira integration creates tickets via a REST API with the appropriate project, issue type and priority, with descriptions that include contributing factors and recommended actions. The email integration uses the Resend API to send HTML-formatted emails to account managers and executives, with templates rendered dynamically based on alert type.

The Decision Agent performs severity classification and rule-based routing. It receives inputs from the Monitoring and Prediction agents, applies business rules to classify severity as Critical (churn > 70% and worsening trend), High (churn > 70% or multiple anomalies), Medium (churn > 60% and worsening trend) or Low (all other cases), and routes decisions to the Action Agent. Fig -4 illustrates the customer-detail view integrated with the Decision Agent's reasoning panel.

4.7 LLM Orchestration with LangChain To manage interactions between the intelligent agents and the underlying LLM API, the LangChain framework [25] is used with Google Gemini as the foundation model. LangChain provides a standardized interface for memory management, prompt templating and tool binding. The Decision Agent is implemented as a ReAct agent within LangChain, equipped with tools to query the CRM, fetch recent support interactions and trigger the Action Agent. The prompt template incorporates dynamic variables – churn probability, CLV, recent anomalies and recent support tickets – to provide context to the LLM and to instruct it to formulate a targeted retention strategy with documented reasoning before acting. This orchestration layer abstracts the complexity of API interactions and ensures that agent responses are parsed robustly into actionable JSON schemas before being dispatched to the Action Execution Layer.

Fig -4: Customer risk profile and LLM-based decision reasoning interface The Action Agent executes workflow automation and triggers: it sends multi-channel notifications, creates Jira tickets for retention or operations teams, sends emails to account managers, triggers automated workflows and logs all execution results. Action mapping is severity based: Critical alerts trigger Slack, Jira and email simultaneously; High triggers Slack and Jira; Medium triggers Slack only; Low-severity events are logged to the database. The Reporting Agent aggregates activity from all agents to generate daily and weekly executive summaries, maintains comprehensive audit trails and delivers reports via email. The Feedback Agent analyzes model performance over time by tracking average prediction confidence, confidence drift between the prediction and decision phases, and action success rates using a 7-day rolling window, and triggers model retraining when performance degrades below configured thresholds.

5. IMPLEMENTATION 5.1 Technology Stack Table -1 summarizes the technology stack used across the three primary layers of the system. Table -1: Technology stack summary

4.5 Agent Communication Protocol Agents communicate through a message bus implemented using database-backed message passing. Each message carries a from agent id, a to agent id, a message type, a JSON payload and an ISO-8601 timestamp. Messages are stored in the agent_communications table, providing full traceability and enabling asynchronous processing.

Impact Factor value: 8.315

Layer Frontend

Component Framework

Frontend Frontend

Language Styling

Frontend

State

Frontend Database

Charts Engine

Database

ORM

Technology Next.js 16 (App Router) TypeScript Tailwind CSS v4, shadcn/ui TanStack React Query Recharts PostgreSQL (Supabase) Supabase Client

ISO 9001:2008 Certified Journal

Page 75

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

Database

Real-time

Database ML Service ML Service

Auth API ML

ML Service ML Service

Time Series LLM

www.irjet.net

(typed) Supabase Realtime Supabase Auth FastAPI (Python) scikit-learn, XGBoost Prophet Google Gemini (LangChain)

Risk_Alerts

+ 50+

trail Active and resolved alerts

5.3 Autonomous Execution The system operates autonomously through scheduled cron jobs. The predictions cron runs every 6 hours: it verifies the cron secret, fetches active customers, runs the prediction pipeline for each customer, and triggers the agent orchestrator's agent chain for any customer whose churn probability exceeds 0.7. The anomaly-detection cron runs every hour, monitoring KPIs and triggering the Monitoring Agent when deviations are detected. The weekly-report cron runs every Monday at 9 AM, generating comprehensive executive summaries and delivering them via email and Slack.

The resulting user interface is a responsive, real-time dashboard displaying key business metrics, recent activities and agent status logs, shown in Fig -5.

5.4 Agent Orchestration Implementation The MultiAgentOrchestrator class coordinates agent execution following the sequential pipeline formalized in Fig -6. Each agent logs its decisions to the database with reasoning, confidence scores and execution metadata. Input: Customer ID c, thresholds θchurn, θact Output: Action execution result or skip D ← FetchCustomerData(c) A ← AnalystAgent.analyze(D) { KPI & trend analysis } F ← ForecastAgent.predict(c, A) { ML predictions } if F.churn_prob > θchurn then δ ← DecisionAgent.evaluate(c, A, F) if δ.shouldAct then ActionAgent.execute(δ, A) ReportingAgent.log(c, A, F, δ) end if end if FeedbackAgent.track(c, F) { Monitor drift }

Fig -5: SIM-OPS real-time operations and KPI monitoring dashboard

5.2 Database Schema The system uses 10 primary tables, summarized in Table -2. Table -2: Database schema overview Table Agents

Rows 6

Agent_Configs

Agent_Decisions

1000+

Agent_Metrics

Agent_Communic ations Workflows Workflow_Runs

500+

Ml_Predictions Activity_Logs

5000+ 10000

3 100+

Fig -6: Autonomous agent chain execution (algorithm)

Purpose Agent metadata and status Agent configuration and thresholds Decision history with reasoning Performance metrics per agent Inter-agent messages

5.5 Deployment The system is deployed on the Vercel Edge Network with Vercel Cron for scheduled tasks, Supabase-hosted PostgreSQL for the database, and the ML service deployed separately on Railway/Render. Production environment variables securely store all API keys.

6. EXPERIMENTAL RESULTS 6.1 Machine Learning Performance

Workflow definitions Workflow execution logs ML prediction results System activity audit

Impact Factor value: 8.315

The churn-prediction model was trained on the IBM Telco Customer Churn dataset [29], comprising 7,043 customers with 21 features spanning demographic,

ISO 9001:2008 Certified Journal

Page 76

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

account and service attributes. Preprocessing included one-hot encoding of categorical variables, median imputation of missing TotalCharges values, and feature engineering to derive tenure-based and behavioral indicators. The dataset was split 80/20 using stratified sampling to preserve the class distribution (73.5% nonchurn, 26.5% churn).

The CLV prediction model achieved a Mean Absolute Error of $127, a Root Mean Squared Error of $189 and an R² score of 0.82. The anomaly-detection model achieved a true-positive rate of 91.3%, a false-positive rate of 4.2% and a detection latency of under 1 second. Fig -8 shows the relative feature importance generated by the XGBoost model, with payment failures, support-ticket count and tenure emerging as the strongest predictors of churn.

Three classifiers were evaluated on the held-out test set to justify the choice of XGBoost as the primary churnprediction model, summarized in Table -3. XGBoost outperformed both baselines across all metrics, with a 3.2% absolute improvement in accuracy over Random Forest and 6.8% over Logistic Regression. The higher recall (88.7%) is particularly important in churn prediction, since missed at-risk customers represent direct revenue loss. Table -3: Churn prediction model comparison Model Logistic Regression Random Forest XGBoost

Accuracy 80.1%

F1 78.3%

AUC 0.84

83.7%

82.1%

0.88

86.9%

86.4%

0.91

Fig -8: Relative feature importance generated by the XGBoost model

6.2 Agent Performance

Table -4: Churn prediction model performance Metric Accuracy Precision Recall F1-Score AUC-ROC Average Confidence

Table -5 summarizes per-agent decision volume, average confidence and response time.

Value 86.94% 84.2% 88.7% 86.4% 0.91 0.89

Table -5: Agent performance metrics Agent Monitoring Prediction Decision Action Reporting Feedback

Decisions 1247 456 892 324 4521 89

Confidenc e 89% 84% 91% 98% 99% 92%

Resp. Time 0.3s 2.1s 0.5s 1.2s 0.2s 5.0s

6.3 System Performance Throughput: 600 predictions per hour, 3,000+ agent decisions per day, 150+ actions executed per day, and an average end-to-end latency of 8.7 seconds. Reliability: 99.7% system uptime, 99.2% successful predictions, 98.9% action-execution success rate, and 99.9% database query success rate. Autonomy: 2 manual interventions per week (down from 20), a 96% autonomous decision rate, a 3.8% false-positive alert rate, and a time-to-action for critical alerts of under 2 minutes.

Fig -7: ROC curve for the XGBoost churn prediction model

Impact Factor value: 8.315

ISO 9001:2008 Certified Journal

Page 77

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

6.4 Business Impact

execution would be necessary. New deployments also face a cold-start problem: the system needs 7–14 days of operation to build sufficient history for accurate forecasting.

Operational efficiency gains include a 90% reduction in manual monitoring, an 85% improvement in response speed to churn signals, an 81% success rate for retention campaigns, and 15 hours saved per week. Based on a projected six-month deployment simulation, estimated cost savings include a 12% improvement in churn reduction, $45K per month in prevented revenue loss, a 35% reduction in operational cost, and a projected ROI of 340% within the first six months.

7.3 Lessons Learned Business-rule thresholds require careful tuning based on industry and company size; a churn threshold of 70% worked well for B2B SaaS but may need adjustment for other domains. Initial deployments generated too many medium-severity alerts, which was addressed through alert aggregation and daily summaries. ML model confidence scores also needed calibration to align with business risk tolerance, which was achieved through confidence thresholds that balance false positives and false negatives.

6.5 Comparison with Baseline SIM-OPS was compared against a baseline manualmonitoring system, summarized in Table -6. Table -6: Comparison with manual baseline Metric Response Time Coverage Accuracy Cost per Alert Human Hours/Week

Manual 4–24 hours 20% 72% $15 20

8. FUTURE WORK

SIM-OPS < 2 minutes 100% 87% $0.50 2

Future iterations will incorporate LSTM networks for time-series churn prediction, NLP-based sentiment analysis of support tickets, and reinforcement learning for optimizing action selection. SHAP values [26] and LIME [27] will be integrated to provide model-agnostic explanations following interpretable machine learning principles [28]. Additional agent types – including Negotiation, Optimization and Simulation agents – will expand autonomous decision-making capabilities. To support larger customer bases (beyond 10,000), message queues (Kafka) for parallel agent execution and Redis for prediction caching are planned. Integration expansion targets CRM platforms (Salesforce, HubSpot), analytics tools (Mixpanel, Amplitude) and data warehouses (Snowflake, BigQuery) to broaden the system's dataingestion capabilities.

7. DISCUSSION 7.1 Key Findings The six-agent architecture proved highly effective for decomposing complex business operations into specialized, manageable components. Agent communication through the message bus provided full traceability while maintaining loose coupling. The system successfully operates 24/7 with minimal human intervention: scheduled cron jobs ensure continuous monitoring, and event-based triggers enable immediate response to critical situations. Every agent decision is logged with detailed reasoning, confidence scores and execution metadata – transparency that is crucial for business stakeholders to trust autonomous systems.

9. CONCLUSIONS This paper presented SIM-OPS, an autonomous multiagent system for intelligent business operations management. The six-agent architecture – comprising Monitoring, Prediction, Decision, Action, Reporting and Feedback agents – demonstrates that complex business operations can be effectively automated through coordinated agent collaboration. Key achievements include 86.94% accuracy in churn prediction with 0.89 average confidence; a 90% reduction in manual intervention while achieving a 96% autonomous decision rate; sub-second response times for critical alerts with multi-channel delivery; full production deployment with 99.7% uptime and 24/7 autonomous operation; and comprehensive audit trails with explainable decisionmaking. The system demonstrates that autonomous AI agents can successfully manage real-world business operations when designed with appropriate specialization, coordination mechanisms and human-oversight

7.2 Limitations While the Feedback Agent monitors model performance, retraining currently requires manual intervention; future work should implement a fully automated retraining pipeline. Setting up integrations with Slack, Jira and email requires configuration and API keys, and a more streamlined onboarding process would improve adoption. The current implementation processes customers sequentially, so for larger customer bases (beyond 10,000) parallel processing and distributed

Impact Factor value: 8.315

ISO 9001:2008 Certified Journal

Page 78

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

capabilities. Future work will focus on enhanced ML capabilities, advanced agent behaviors, scalability improvements and broader integration with enterprise systems.

12) M. Dumas, M. La Rosa, J. Mendling, and H. A. Reijers, Fundamentals of Business Process Management, 2nd ed. Berlin, Germany: Springer, 2018. 13) S. Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” in International Conference on Learning Representations (ICLR), 2023.

ACKNOWLEDGEMENT The authors would like to thank the open-source community for the excellent tools and libraries that made this work possible, including Next.js, FastAPI, Supabase and the Python machine-learning ecosystem.

14) S. J. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed. Hoboken, NJ: Pearson, 2020.

REFERENCES

15) J. Ferber, Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence. Reading, MA: Addison-Wesley, 1999.

1) M. Wooldridge, An Introduction to MultiAgent Systems, 2nd ed. Chichester, UK: John Wiley & Sons, 2009.

16) P. Stone and M. Veloso, “Multiagent systems: A survey from a machine learning perspective,” Autonomous Robots, vol. 8, no. 3, pp. 345–383, 2000.

2) R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA: MIT Press, 2018. 3) P. Dayan and G. E. Hinton, “Feudal reinforcement learning,” in Advances in Neural Information Processing Systems, 1992, pp. 271–278.

17) L. Busoniu, R. Babuska, and B. De Schutter, “A comprehensive survey of multiagent reinforcement learning,” IEEE Trans. Syst., Man, Cybern. C, vol. 38, no. 2, pp. 156–172, 2008.

4) G. Shmueli and P. C. Koppius, “Predictive analytics in information systems research,” MIS Quarterly, vol. 35, no. 3, pp. 553–572, 2011.

18) J. S. Park et al., “Generative agents: Interactive simulacra of human behavior,” in Proc. 36th ACM UIST, 2023, pp. 1–22.

5) A. Bifet, R. Gavalda, G. Holmes, and B. Pfahringer, Machine Learning for Data Streams. Cambridge, MA: MIT Press, 2018.

19) Z. Xi et al., “The rise and potential of large language model based agents: A survey,” arXiv preprint arXiv:2309.07864, 2023.

6) X. He, K. Zhao, and X. Chu, “AutoML: A survey of the state-of-the-art,” Knowledge-Based Systems, vol. 212, p. 106622, 2021.

20) E. W. T. Ngai, L. Xiu, and D. C. K. Chau, “Application of data mining techniques in customer relationship management: A literature review and classification,” Expert Syst. Appl., vol. 36, no. 2, pp. 2592–2602, 2009.

7) W. Verbeke, D. Martens, C. Mues, and B. Baesens, “Building comprehensible customer churn prediction models with advanced rule induction techniques,” Expert Systems with Applications, vol. 38, no. 3, pp. 2354–2364, 2011.

21) L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. 22) T. Vafeiadis, K. I. Diamantaras, G. Sarigiannidis, and K. Ch. Chatzisavvas, “A comparison of machine learning techniques for customer churn prediction,” Simulation Modelling Practice and Theory, vol. 55, pp. 1–9, 2015.

8) Y. Xie, X. Li, E. W. T. Ngai, and W. Ying, “Customer churn prediction using improved balanced random forests,” Expert Systems with Applications, vol. 36, no. 3, pp. 5445–5449, 2009.

23) F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in Proc. 8th IEEE Int. Conf. Data Mining, 2008, pp. 413–422.

9) T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 785–794.

24) S. J. Taylor and B. Letham, “Forecasting at scale,” The American Statistician, vol. 72, no. 1, pp. 37– 45, 2018.

10) Apache Airflow. (2015). Apache Airflow Documentation. [Online]. Available: https://airflow.apache.org/ 11) Temporal Technologies. (2019). Workflow Engine. [Online]. https://temporal.io/

25) LangChain, Inc. (2023). LangChain: Building applications with LLMs through composability. [Online]. Available: https://github.com/langchain-ai/langchain

Temporal Available:

Impact Factor value: 8.315

ISO 9001:2008 Certified Journal

Page 79

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 13 Issue: 08 | Aug 2026

p-ISSN: 2395-0072

www.irjet.net

26) S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Advances in Neural Information Processing Systems, vol. 30, 2017. 27) M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should I trust you? Explaining the predictions of any classifier,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 1135–1144. 28) C. Molnar, Interpretable Lulu.com, 2020.

Machine

Learning.

29) IBM Developer. (2018). Telco Customer Churn Dataset. [Online]. Available: https://www.kaggle.com/datasets/blastchar/telc o-customer-churn

Impact Factor value: 8.315

ISO 9001:2008 Certified Journal

Page 80