🤖 SLA Risk Prediction — IT Ticket Management (ZF Automotive)

100% original project — independently designed and developed by Silas Luiz Bom Fim Applied AI for IT service management in an automotive manufacturing environment.

---

📌 Overview

This project develops a hybrid AI model combining Decision Trees and LSTM Neural Networks to predict SLA breach risk in IT ticket management at ZF Automotive Brasil (Limeira unit).

The entire solution — from problem definition to data architecture and model implementation — was conceived and built independently, using a synthetic sandbox dataset modeled after real operational patterns to protect corporate confidentiality.

---

🎯 Problem Statement

The IT support team at ZF Automotive operates reactively, processing tickets by arrival order and nominal SLA priority. This generates three critical bottlenecks:

Bottleneck	Description
Triage Blindspot	Tickets that appear simple at opening but carry hidden technical complexity
Accumulation Inertia	Backlog from previous days creates "operational fatigue" invisible to simple statistics
Capacity Waste	High responsiveness to low-priority (P4) tickets reduces senior specialist availability for critical incidents

Business goal: Shift from reactive to predictive IT management, anticipating SLA breaches before they occur.

---

🏗️ Solution Architecture

A hybrid two-model approach:

Ticket Opens → Decision Tree (immediate triage) → LSTM (weekly risk forecast) → Action Plan

🌳 Decision Tree — "The Gatekeeper"

• Role: Immediate triage at ticket opening
• Output: Decision rules (IF complexity ≥ 3 AND no knowledge base → Senior escalation)
• Result: 97.3% triage accuracy

🧠 LSTM Neural Network — "The Manager"

• Role: Weekly SLA risk forecasting based on workload history
• Output: Risk probability for the next operational period
• Result: 84.2% accuracy after 50 training epochs
• Key insight: Learns temporal dependencies — yesterday's backlog affects today's risk

---

📊 Dataset — Sandbox Environment

⚠️ Important: All data in this project is synthetic, generated to replicate the statistical patterns of real ZF Automotive operational data without exposing any confidential corporate information.

Coverage: H2 2025 (July 1 – December 31, 2025) Size: 184 daily records

Feature	Type	Description
Data	Temporal	Date — essential for LSTM temporal dependency learning
Vol_P1 to Vol_P4	Integer	Ticket volume by SLA priority level (P1 = critical, P4 = low)
Complexidade_Cat	Categorical	Technical complexity scale (1 = simple, 4 = out of scope)
Tempo_Ocioso_Min	Continuous	Minutes from ticket opening to first analyst action (responsiveness target)
Tempo_Atuacao_Min	Continuous	Effective working time on the ticket
BC_pronta	Binary	Knowledge base solution already available (1 = yes)
BC_possivel	Binary	Knowledge base solution could be created (1 = yes)
Risco_SLA	Binary	Target variable — SLA breach risk (1: Risk / 0: Normal)

---

📈 Results

Model Performance

Model	Metric	Value
Decision Tree	Triage Accuracy	97.3%
LSTM Neural Network	Predictive Accuracy	84.2%

Business Impact — Hybrid Strategy Simulation

Scenario	Days at Risk (Semester)	Change
Without AI — baseline	23 days	—
Knowledge Base + Automation	9 days	−60.9%

The combined strategy of structured knowledge bases + process automation reduced SLA breach risk days from 23 to 9 in a single semester — a 60.9% improvement in operational reliability.

---

🔄 Business Logic — Triage Decision Engine

def definir_plano(row):
    if row['BC pronta'] == 1:
        return 'Suggest Solution (Ready KB)'     # Reduces resolution time
    elif row['BC possivel'] == 1:
        return 'Action: Create Documentation'    # Backlog for knowledge capture
    elif row['Complexidade_Cat'] >= 3:
        return 'Senior Analysis Required'        # New complex cases
    else:
        return 'Normal Flow'

---

🛠️ Tech Stack

pandas · numpy · matplotlib        # Data handling & visualization
scikit-learn                       # Decision Tree, preprocessing
tensorflow / keras                 # LSTM Neural Network
imbalanced-learn                   # Class balancing

---

🚀 How to Run

1. Open the notebook in Google Colab (recommended)
2. Upload the sandbox CSV file when prompted
3. Run all cells sequentially

# If running locally
pip install pandas numpy matplotlib scikit-learn tensorflow imbalanced-learn

---

💡 Future Work

• Integration with RPA tools for autonomous resolution of P4 tickets
• Real-time dashboard powered by LSTM for dynamic staffing during predicted peak periods
• Expansion of the knowledge base to push risk days below 9 per semester

---

👤 Author

Silas Luiz Bom Fim Data Engineer · ML & AI Developer · UFABC

• 💼 LinkedIn
• 🐙 GitHub
• 📧 silas.luiz.96@gmail.com