Levon Y.

Grid Dynamics is a digital engineering company providing AI, analytics, and cloud solutions to Fortune 500 enterprises. Engaged on an Apple project.

• Conducted safety evaluations of summarization systems for app store reviews and music search queries.
• Generated synthetic datasets and designed prompt-based classification pipelines using various large language models (LLMs).
• Identified unsafe model outputs and contributed to improving model alignment with established safety standards.
• Optimized prompt engineering strategies to improve classification accuracy across multiple content categories.

Simple Life is a health and wellness platform focused on personalized weight loss and user retention through data-driven task assignment.

• Developed a personalized task assignment system to improve user weight loss outcomes and increase retention rates.
• Led the end-to-end release of recommendation modules to all production users.
• Built data pipelines and analytical models using Python, SQL, Amazon SageMaker, and Snowflake.
• Applied retention analysis using Retentioneering to identify churn patterns and optimize user engagement flows.

Portmind is a technology company delivering ML-powered solutions for evidence prediction, document intelligence, and industrial automation.

• Led a team in improving evidence prediction through similar description identification and uplift classification, recovering 10% of previously unresolved cases.
• Leveraged large language models (LLMs), including OpenAI models and Zephyr 7B, for feature engineering, automating 80% of financial fraud detection workflows.
• Managed a truck and container number detection project, increasing the size of the domain-specific dataset by 75% through web scraping and synthetic data generation.
• Built object detection and character recognition models using PyTorch and YOLO, achieving 96% object detection accuracy and 80% character recognition accuracy.
• Collaborated on a patent classification solution for Philip Morris International, designing a multi-label classification pipeline using Scikit-learn and LightGBM.
• Assessed label quality through in-depth annotation analysis, identifying an estimated 15% mislabeling rate.
• Created interactive data visualizations using Streamlit and deployed machine learning solutions with Docker.

Webb Fontaine is a technology company specializing in AI-driven customs and trade facilitation solutions for governments worldwide.

• Achieved a 90% automated HS code classification rate by training and optimizing FastText and BERT models.
• Trained a transformer-based spell checker that improved text preprocessing quality for noisy descriptions by 20%.
• Collaborated on the development of fraud detection models for invoices and shipping manifests, improving customs tax processes and security measures.
• Designed and implemented data processing and machine learning training pipelines using Scikit-learn, CatBoost, and LightGBM.
• Developed a clearance time forecasting model using the Holt-Winters forecasting algorithm.
• Built a fake invoice detection system based on document layout analysis using OpenCV.