Elif Pulukcu

› Computer Vision System for Product Attribute Recognition and Outfit Recommendation

› Natural-Language Database Query System (text-to-SQL RAG)

Letting anyone query the database in plain English, no SQL required.

• Built a Retrieval-Augmented Generation (RAG) text-to-SQL system that turns natural-language questions into SQL queries over a retail database, automating data access and reporting.
• Retrieved relevant schema elements (tables, columns, and descriptions) to ground query generation and improve accuracy.
• Used few-shot prompting with curated question-SQL pairs tailored to domain-specific terminology.
• Executed generated queries and returned results as both structured outputs and natural-language summaries, removing manual reporting workflows.

› Retail Data Analytics & Reporting

Turning raw retail data into reports and dashboards business teams actually use.

• Supported business teams with data-driven reporting and analysis in a retail environment, working primarily with Google BigQuery and Oracle databases.
• Developed monthly SQL-based reports to track key business and sales metrics.
• Built and maintained Qlik Sense dashboards for ongoing monitoring of business and customer performance.
• Conducted top-customer analysis to support customer-focused decisions, and resolved ad-hoc data requests from business teams with custom SQL queries.

Worked across machine learning, industrial IoT and data engineering for predictive maintenance and production traceability, moving real-time factory data all the way from shop-floor devices to cloud deployment on AWS and Docker.

› Industrial IoT Predictive Maintenance System

Catching machine failures before they happen, from raw sensor streams to live alerts.

• Built a Variational Autoencoder (VAE)-based anomaly detection system on real-time industrial IoT sensor data, covering preprocessing, training, validation and metric evaluation.
• Migrated alerting from rule-based to ML-based, improving anomaly detection coverage.
• Set up AWS Kinesis Data Streams for real-time processing and AWS Timestream for storage, with error-handling for failed messages.
• Deployed the inference API with Docker and set up monitoring dashboards and system tests.

› Production Traceability for Washing Machine Manufacturing

Following every washing machine down the line to trace exactly where things go wrong.

• Built a traceability system to monitor critical process parameters across washing machine production.
• Collected and preprocessed production data, integrated IoT devices for real-time tracking, and designed pipelines for data collection and storage.
• Implemented root-cause analysis workflows to pinpoint error sources on the production lines.
• Documented workflows and reported on system performance and findings.

Worked mostly with automotive data, building computer vision and machine learning solutions end to end and training models on AWS.

› Automotive Dashboard Warning Light Recognition

Teaching a model to read the cryptic symbols on a car's dashboard.

• Built CNN-based models to detect and classify dashboard warning indicators from customer-submitted images, owning the pipeline from preprocessing and dataset preparation to training and evaluation.
• Connected predictions to action by mapping each detected warning to the right maintenance and repair service request.
• Benchmarked localization and classification with IoU and mAP@0.5 / mAP@0.5:0.95.

› Automated Labor Planning

Helping service centers staff the right number of technicians on the right days.

• Built a workforce-planning solution pairing labor-demand forecasting with linear-programming staff scheduling.
• Mined historical service operations and appointment data to estimate workforce needs and support capacity planning.
• Optimized technician allocation across service centers, cutting customer waiting times and improving efficiency.

› TV Ratings Forecasting

Predicting who’s tuning in, episode by episode.

• Built an ensemble combining ARIMA for time-series analysis and XGBoost for machine learning to forecast episode-level TV ratings.
• Engineered features from broadcast metadata, significantly reducing forecast error (measured with RMSE and MAE).