Hire Harshith P. - AI Developer

Engineered 100+ predictive featuresfrom 1M daily IoT sensor readings using Python (Pandas, NumPy) and Apache Spark (PySpark), applying FFT for vibration analysis and statistical aggregations for time-series data. -Developed and deployed an ensemble of supervised learning models including XGBoost and LightGBM, achieving 88% recall in predicting equipment failures 24-72 hours in advance by leveraging SMOTE to handle class imbalance. -Built unsupervised learning modelslike Isolation Forest and Autoencoders (PyTorch) for real-time anomaly detection, achieving 92% accuracy in identifying abnormal sensor patterns indicating early equipment degradation. -Implemented Computer Vision models using PyTorch CNNs (ResNet-50) for visual inspection of equipment, achieving 94% defect detection accuracy via transfer learning on 100,000 labeled images deployed on edge devices. -Designed a Reinforcement Learning agent using Deep Q-Networks (PyTorch) to learn optimal maintenance policies, balancing costs and uptime, which reduced maintenance costs by 35% and prevented 70% of unplanned downtime. -Orchestrated end-to-end ETL/ELT pipelines using Apache Airflow and stored processed data in AWS S3 and PostgreSQL, enabling efficient feature storage and retrieval for model training and inference. -Established a robust MLOps lifecycle using MLflow for experiment tracking and Docker/Jenkins for CI/CD, automating model retraining and deployment on AWS SageMaker while monitoring for data drift with AWS CloudWatch.

* Delivered ETL and ELT pipelines consolidating heterogeneous client data into cloud warehouses using Python, SQL, PySpark, and Apache Airflow, reducing data-quality incidents by 40%. * Constructed ensemble failure-prediction models using XGBoost and LightGBM with 85% recall on 48-hour horizons, enabling proactive maintenance planning for industrial clients. * Built unsupervised anomaly detection layers with Isolation Forest and autoencoders, integrating outputs via REST APIs to enhance operational incident management workflows. * Created interactive Tableau and Power BI dashboards for executive KPI tracking, standardizing metric definitions and increasing dashboard resilience to team changes. * Improved retail demand-forecasting accuracy from 14% to 8% MAPE by blending classical time-series and LSTM models, supporting leaner inventory management. * Integrated SHAP interpretability and partial-dependence visualizations into client models to facilitate transparent decisionmaking in regulated environments. * Instituted MLflow for experiment tracking and developed Dockerized microservices for scalable model deployment with structured logging and latency monitoring.