Eduard Mihranyan, Developer in Yerevan, Armenia
Eduard is available for hire
Hire Eduard

Eduard Mihranyan

Verified Expert  in Engineering

Machine Learning Developer

Location
Yerevan, Armenia
Toptal Member Since
January 10, 2022

Eduard is an experienced data scientist with a demonstrated history of working in IT companies and the banking industry. Having more than seven years of experience in the industry, he has proved his proficiency in providing high-quality end-to-end solutions that significantly improve company KPIs. His recent projects are in the Generative AI field, specifically LLMs and Text2Image models. Eduard is a problem solver. He continuously improves his arsenal by learning and always staying up to date.

Portfolio

Plat.AI
Text to Image, Analytics, Machine Learning, Deep Learning, Finance, Python...
The Estee Lauder Companies Inc.
Artificial Intelligence (AI), Google AI Platform, Recommendation Systems...
PicsArt
Machine Learning, Deep Learning, Data Analysis, Engineering, CI/CD Pipelines...

Experience

Availability

Part-time

Preferred Environment

PyCharm, Slack, GitHub, Amazon Web Services (AWS), ARIMA, Forecasting, JupyterLab

The most amazing...

...thing I've developed is an in-house recommendation engine for one of the biggest photo editing companies.

Work Experience

Senior Machine Learning Engineer

2022 - PRESENT
Plat.AI
  • Created an object replacement model, which finds the described object in the image and replaces it with the provided image of the new object, keeping everything else the same.
  • Fine-tuned an instruction-based LLM to solve a specific summarization task for the company data.
  • Developed a reinforcement learning environment to explore and distribute resources optimally, maximizing the company's profit. The project had the highest priority and showed significant improvements.
  • Designed an insurance scoring model, which estimates the probability of accidents and calculates the insurance rate based on the risk.
  • Built a customer time series prediction model to be used in the banking industry for improving their scoring models.
Technologies: Text to Image, Analytics, Machine Learning, Deep Learning, Finance, Python, OpenAI GPT-3 API, Language Models, Reinforcement Learning, Amazon Web Services (AWS), Generative Pre-trained Transformer 3 (GPT-3), Generative Pre-trained Transformers (GPT), OpenAI GPT-4 API, Predictive Modeling, Pricing Models, Git, Docker, Data Scientist, ChatGPT, OpenAI, Azure, Azure SQL Data Warehouse, Dedicated SQL Pool (formerly SQL DW), Google Cloud Platform (GCP), SQL, Large Language Models (LLMs), Data Processing, Kubernetes

Google Recommendation AI Specialist

2022 - 2022
The Estee Lauder Companies Inc.
  • Developed an automated recommendation pipeline that collects and preprocesses data, trains a model, and serves as an API endpoint to request and get recommendations.
  • Designed a comprehensive A/B test for checking the recommendation model's online performance.
  • Created and trained in-house recommendation models—user-based and item-based—which may be used to replace the current solution in the future.
Technologies: Artificial Intelligence (AI), Google AI Platform, Recommendation Systems, Google, Python, Machine Learning, APIs, Google BigQuery, AI Design, Neural Networks, Forecasting, MySQL, Data Pipelines, Snowflake, Model Development, Data Analytics, Data Extraction, Data Engineering, Git, Data Scientist, Google Cloud Platform (GCP), Google Analytics, BigQuery, Data Processing

Full-stack Machine Learning Scientist

2019 - 2022
PicsArt
  • Built an in-house text-to-image generation model based on Stable Diffusion, which works live, and an integrated fine-tuning mechanism so the generator could create and modify users' pictures, e.g., create a superhero with your face.
  • Developed recommender system models for sticker recommendations to a user. The first model was based on user preferences and has shown more than 300% growth in sticker usage in each touchpoint when rolled out.
  • Grew premium content usage by 150% and saw a significant increase in subscription metrics after creating a model that considers the visual match between a photo and sticker used for cold-start users and premium stickers.
  • Boosted the marketing team's campaign success rates in targeting audiences by designing a model that captures user preferences toward PicsArt tools and content and divides them into meaningful segments.
  • Created an anomaly detection algorithm that captures anomalous spikes in app crashes and reports them to the development team, improving the team's process for fixing bugs.
  • Increased the company's revenue significantly by developing an ads optimization model that showed users ads and subscription offer screens.
  • Conducted a global company analysis for the CEO to find optimal ways of development, which changed the company's growth direction and resulted in a new department organized with the highest priority projects.
  • Managed a team of three members working on different projects and encouraged their personal growth.
Technologies: Machine Learning, Deep Learning, Data Analysis, Engineering, CI/CD Pipelines, Torch, Reinforcement Learning, Statistics, Research, Python, PySpark, Pandas, Scikit-learn, Gensim, NumPy, Artificial Intelligence (AI), Recommendation Systems, GitHub, Data Science, Google AI Platform, Mathematics, Analytics, Time Series, R, APIs, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), AI Design, Time Series Analysis, Amazon Web Services (AWS), Neural Networks, Artificial Neural Networks (ANN), TensorFlow, Image Generation, ARIMA, ARIMA Models, Forecasting, LSTM, Natural Language Understanding (NLU), PyTorch, JupyterLab, MySQL, Random Forests, XGBoost, Data Pipelines, Linear Regression, Snowflake, Model Development, Keras, Computer Vision, BERT, Data Analytics, Data Reporting, Language Models, Data Extraction, Data Engineering, Predictive Modeling, Predictive Analytics, Databricks, Azure Databricks, Data-driven Marketing, Pricing Models, Git, Docker, Algorithms, Data Scientist, PostgreSQL, Plotly, Tableau, Data Manipulation, Data Modeling, OpenAI, Azure, Azure SQL Data Warehouse, Dedicated SQL Pool (formerly SQL DW), Google Cloud Platform (GCP), Statistical Data Analysis, SQL, Google Analytics, BigQuery, Transformers, Data Processing, Kubernetes

Machine Learning Scientist

2018 - 2019
Ameriabank
  • Developed a credit risk model to predict defaults of retained clients that currently works as the primary method for landing retail credits. The result is a significant increase in the company's profit.
  • Created a model to predict the risk of corporate clients as the primary method for landing corporate credits, which resulted in a significant increase in the company's profit.
  • Developed a credit portfolio optimization model to increase the profitability of the current portfolio.
Technologies: Machine Learning, Deep Learning, Finance, Data Analysis, Optimization, Python, Pandas, Torch, Scikit-learn, Gensim, NumPy, Artificial Intelligence (AI), GitHub, Data Science, Mathematics, Analytics, Time Series, R, APIs, AI Design, Architecture, Time Series Analysis, Amazon Web Services (AWS), Microsoft Power BI, Neural Networks, Artificial Neural Networks (ANN), TensorFlow, AI Programming, ARIMA, ARIMA Models, Forecasting, LSTM, Natural Language Understanding (NLU), PyTorch, JupyterLab, MySQL, Random Forests, XGBoost, Data Pipelines, Linear Regression, Model Development, Keras, Computer Vision, BERT, Data Analytics, Data Reporting, Language Models, Data Extraction, Predictive Modeling, Predictive Analytics, Data-driven Marketing, Git, Algorithms, Data Scientist, PostgreSQL, Tableau, Data Manipulation, Financial Analysis, Statistical Data Analysis, SQL

Machine Learning Engineer

2017 - 2018
BetConstruct
  • Improved the previous prototype for modeling the final score of soccer matches.
  • Developed a model that predicted corner kicks and the number of yellow or red cards in a soccer match.
  • Created a model to predict the final result of a basketball match—win, loss, or tie.
Technologies: Machine Learning, Deep Learning, Gaming, Python, Pandas, Data Analysis, Scikit-learn, NumPy, Artificial Intelligence (AI), GitHub, Data Science, Mathematics, Analytics, Time Series, R, AI Design, Microsoft Power BI, Neural Networks, Artificial Neural Networks (ANN), TensorFlow, AI Programming, Forecasting, LSTM, Natural Language Understanding (NLU), PyTorch, JupyterLab, MySQL, Random Forests, XGBoost, Linear Regression, Model Development, Keras, Computer Vision, BERT, Data Analytics, Data Reporting, Data Extraction, Predictive Modeling, Algorithms, Data Scientist, Google Analytics

Data Scientist

2016 - 2017
TeamViewer Germany
  • Collaborated with the growth hackers team and analyzed the main direction of the company's growth; found out optimal bundling of current products and optimal prices for sales.
  • Conducted churn analysis based on comments from past users, which led to the main reasons for client churn (text analysis).
  • Created analytic dashboards for monitoring company-wide metrics.
Technologies: Machine Learning, Statistics, Data Analysis, SQL, R, Pandas, Scikit-learn, NumPy, Data Science, Mathematics, Analytics, Time Series, AI Design, Time Series Analysis, Microsoft Power BI, Neural Networks, Artificial Neural Networks (ANN), AI Programming, ARIMA Models, Forecasting, JupyterLab, Random Forests, XGBoost, Linear Regression, BERT, Data Analytics, Data Reporting, Predictive Analytics, Data-driven Marketing, Algorithms, Data Scientist, Tableau, Google Analytics

Text to Image Generator

A Stable Diffusion-based model is trained to create images using Gaussian noise while taking user-provided guidance in the form of images and text. By giving the model a user's photo and accompanying text instructions, the generator can produce a new image that modifies the original photo according to the given instructions. For example, it can create a superhero image with the user's face and position the superhero near the Eiffel Tower.

Photo-based Recommender System

The model captured the best fit between photos and stickers based on previously selected, excellent quality photo edits. It used a metric learning approach to obtain the distance between a photo and sticker based on their visual characteristics.

Preference-based Recommender System

The model captured user-to-item and item-to-item interactions to predict and recommend new items to the current users. We trained the model on vast data of PicsArt users and their usage of stickers. Currently, the model is in production on all platforms and touchpoints.

Ads Optimization

In this project, the model captured user preferences and interests to optimize ads in the application and show subscription offer screens based on the tools and content each user tends to like the most.

Credit Default Prediction

The project dealt with predicting the probability of default on each loan and determining the creditworthiness of each client. The amount and rate of the landing were decided for each loan, or if the probability of default was high, the loan request was denied.

Marketing Campaign Optimization

My role in this project was to create a machine learning model that discovered users' interests toward features of the application and divided all users into subgroups with similar interests. The marketing team used this to run the targeted marketing campaigns on the smaller groups with higher conversion rates in key metrics.

Time Series Anomaly Detection

The model captured anomalous changes in the behavior of hourly count data crashes, analyzing the history of time series data. When the anomaly was detected, it alerted the development team to fix issues.

Statistical Guarantees for Conditional Generative Models

An applied research project was performed under the supervision of professor Arnak Dalalyan from ENSAE Paris. We introduced a convenient framework for studying conditional (adversarial) generative models from a statistical perspective. It consisted of modeling a generative device as a smooth transformation of a unit hypercube with a dimension much smaller than ambient space and measuring the quality of the generative model through integral probability metrics.

Aesthetic Predictor

A model is trained to predict whether the given photo can be considered aesthetic. The main difference between existing state-of-the-art models is that it captures company-specific standards that do not appear in open source datasets. It is also used as a content quality estimator, which helps increase the performances of other models relying on the quality score.
2019 - 2021

Master's Degree in Data Science

Yerevan State University - Yerevan, Armenia

Libraries/APIs

PySpark, LSTM, PyTorch, XGBoost, NumPy, Pandas, Scikit-learn, TensorFlow, Keras, Spark ML

Tools

Google Analytics, Gensim, GitHub, Git, Tableau, BigQuery, Google AI Platform, Microsoft Power BI, ChatGPT, Plotly

Languages

SQL, Python, R, Snowflake

Paradigms

Data Science

Platforms

Databricks, Amazon Web Services (AWS), Google Cloud Platform (GCP), Kubernetes, Docker, Azure, Azure SQL Data Warehouse, Dedicated SQL Pool (formerly SQL DW)

Storage

MySQL, PostgreSQL, Data Pipelines

Frameworks

Spark

Other

Machine Learning, Deep Learning, Data Analysis, Statistics, Artificial Intelligence (AI), Recommendation Systems, Mathematics, Analytics, Natural Language Processing (NLP), AI Design, Neural Networks, Artificial Neural Networks (ANN), AI Programming, ARIMA, ARIMA Models, Forecasting, JupyterLab, Random Forests, Linear Regression, Model Development, BERT, Data Analytics, Data Reporting, Generative Pre-trained Transformers (GPT), Data Extraction, Predictive Modeling, Probability Theory, Predictive Analytics, Data Scientist, Data Modeling, Statistical Data Analysis, Transformers, Data Processing, Engineering, Torch, Reinforcement Learning, Research, Finance, Optimization, Time Series Analysis, Time Series, Google BigQuery, Architecture, Financial Modeling, SARIMA, Natural Language Understanding (NLU), Computer Vision, Generative Adversarial Networks (GANs), Language Models, Word Embedding, OpenAI GPT-4 API, Azure Databricks, Data-driven Marketing, Pricing Models, Algorithms, Data Manipulation, OpenAI, Generative AI, Large Language Models (LLMs), CI/CD Pipelines, Gaming, APIs, Image Generation, Google, Generative Systems, Diffusion Models, Generative Artificial Intelligence (GenAI), Text to Image, Images, Clips, Aesthetics, Content, Data Engineering, OpenAI GPT-3 API, Generative Pre-trained Transformer 3 (GPT-3), Financial Analysis

Collaboration That Works

How to Work with Toptal

Toptal matches you directly with global industry experts from our network in hours—not weeks or months.

1

Share your needs

Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2

Choose your talent

Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3

Start your risk-free talent trial

Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.

Top talent is in high demand.

Start hiring