Back to projects

IFS Türkiye R&D internship project2026End-to-end application development

AI-Powered Predictive Maintenance & ERP Decision Support

An ERP-oriented decision-support system that turns sensor data into traceable maintenance decisions through failure-risk prediction, SHAP explanations, machine criticality, inventory context, and human approval.

  • AI4I 2020 · approximately 10,000 synthetic records
  • SHAP explanations and saved evaluations
  • Work orders after human approval with immutable audit
SensorRisk + SHAPPriorityHuman approval
Conceptual flow from sensor signal to explainable risk and human approval. Sensor inputs → ML analysis → Decision context → Human approval.
Role

End-to-end application development

Context

IFS Türkiye R&D internship project · 2026

Core stack

Python · Django REST · React · PostgreSQL

Scale

AI4I 2020 · approximately 10,000 synthetic records

Overview

Rather than stopping at a failure score, the project builds an ERP-oriented workflow that turns predictions into explainable, traceable maintenance decisions. Model output is considered together with machine criticality, parts status, and human approval.

Problem and goal

A risk probability alone is not enough for maintenance teams. The influential signals, operational criticality, parts status, and the person approving the decision all need to be visible in one workflow.

My role

  • Developed feature engineering and the Random Forest prediction flow for sensor data.
  • Connected SHAP explanations to saved evaluations and prediction details.
  • Combined dashboard, work-order, parts/inventory, and role-based administration flows in one application.

Solution

  • Ambient/process temperature, rotational speed, torque, tool wear, and product type were normalized into a shared analysis contract.
  • Predictions were presented with influential SHAP features, separating quick analysis from saved evaluations.
  • Maintenance priority, work-order, and inventory context were captured in a decision record while final approval remained human-led.

Architecture

The system’s main components and data/decision flow are separated as follows.

System overview
01Sensor inputsTemperature, speed, torque, wear
02ML analysisFeatures · Random Forest · SHAP
03Decision contextCriticality · parts · priority
04Human approvalWork order and audit trail

Tech stack

Python · scikit-learn · SHAP

Feature engineering, risk prediction, and explanation

Django REST Framework

API, roles, and decision records

React · TypeScript

Analysis and administration interfaces

PostgreSQL · Docker

Persistent data and runtime environment

Implementation decisions

  • Avoid evaluating a rare failure class with accuracy alone; prefer precision, recall, and PR-AUC where reports support them.
  • Prevent target leakage by excluding fields that can indirectly reveal the outcome from training inputs.
  • Treat model explanation as part of the decision record rather than a decorative add-on.

Results and evaluation

  • Delivered a working application flow spanning analysis, prediction history, record detail, work orders, replay, and administration.
  • Made model risk and operational context visible and auditable on the same decision surface.

Challenges and solutions

  • Imbalanced outcomes → an evaluation approach that avoids majority-class accuracy masking useful behavior.
  • Gap between prediction and maintenance action → a workflow combining criticality, inventory, and human approval.

What I learned

The project reinforced that a production-minded ML feature is more than a model: data contracts, explainability, human approval, and traceable records must be designed together.

Next projectThreatGraph AI: Cyber Threat Analysis & Risk Prediction