Research
Publications
- Training Noise Sensitivity Analysis of Default and Tuned Classifiers for NIR Spectroscopic Detection of Pinto Bean Adulteration
Discover Food2026
- Smart Bean Analysis: Rapid Detection with Spectral Data and Deep Learning
Acta Technologica Agriculturae 29(2), 89-972026
- SDPERL: A Framework for Software Defect Prediction Using Ensemble Feature Extraction and Reinforcement Learning
arXiv preprint arXiv:2412.079272024
- Counterfeit Detection of Iranian Black Tea Using Image Processing and Deep Learning Based on Patched and Unpatched Images
Horticulturae 10(7), 6652024
Experience
- MyKolbeh2025 to 2026AI/ML Engineer
Built a GPT-4o photo-and-text valuation model that cut pricing error (MAPE) from 19% to 14%, about a 26% relative drop. It was fed by an XGBoost pricing engine with seasonal and lag features and a GPT-4o-mini agentic crawler that auto-collected competitor listings.
- HABIBI Lab, Sharif University2024 to 2025Research Assistant
Combined ensemble feature extraction with a PPO agent for file-level software defect prediction, gaining +6.25% F1 over baselines on the PROMISE dataset, and co-authored the SDPERL preprint.
- RIML Lab, Sharif University2023 to 2024Research Assistant
Trained a patch-based MobileNetV3 classifier to 95% accuracy for counterfeit black-tea detection, the best of five architectures tested against RegNetY, EfficientNetV2, ShuffleNetV2 and SwinV2T. I also benchmarked classical models on VIS-NIR spectra under rising training noise, where ridge held above 0.95 accuracy while the ensembles degraded sharply, and used SHAP to trace pinto-bean adulteration to the 1400 to 1500 nm water band. Three papers came out of this work, in Horticulturae, Acta Technologica Agriculturae and Discover Food.
- RADIAN Lab, Sharif University2023 to 2024Research Assistant
Designed synthetic-data pipelines (GANs, VAEs, Gaussian mixtures) for WiFi-CSI crowd-size estimation across multiple rooms, raising accuracy from 30% to 43% on unseen environments.
- Dezpart Gohar Kimia Sepehr2024IoT Intern
Instrumented a zinc-sulfate production line with a networked IoT sensor array and trained time-series models to forecast throughput bottlenecks hours ahead, surfaced in a real-time dashboard whose threshold alerts cut unplanned downtime by about 15%.