Igor Stankevich

• Implemented framework (Python, pandas) that helped to simplify modeling in commodities and rewired existing models via common interface
• Developed complex modeling graphs shared caching strategies (MongoDB/Arctic, Redis, AWS S3) that cut down model run times and enabled quicker collaboration and scenarios exploration between analysts
• Provided support for scenarios management (MongoDB) and execution on a cloud (AWS farm + S3)
• Built an extensive test suite, covering more than 300 functional, integration and performance requirements
• Deployed detailed and intuitive documentation, along with use cases description into centralized system
• Worked with head of directional analysts to design a concept and iterate on core aspects of the framework

Developed facial recognition system core with 99.8% match rate (Chainer/Caffe, Tornado, ReactJS), that replaced legacy system with 75% accuracy, and became part of a corporate facial recognition cloud.
Development steps:
1. Detailed research on most recent technologies available
2. Core model coding based on academia papers, and training on publicly available dataset
3. Proof of concept (Flask web server + Angular UI): recognizing people on video
4. Core library, trained on proprietary faces database (50 mln faces), with tests and documentation
5. Cloud integration API and scaling

Developed license plates recognition system using proprietary license plates database (Python, OpenCV, dlib, Tornado, AngularJS), that was integrated into federal highway project on 40 toll lanes.
Development workflow:
1. Solution design: detection approach, manual features, template segmentation and data aggregation approaches
2. Manual and semi-manual labeling with proprietary and public tools, on a private data set
3. Proof of concept implementation: Tornado server + Angular UI, used proprietary online video (brand new real time data)
4. Optimized application for low-performance lane controllers
5. Deployment

Computer vision system to detect and classify vehicles into 4 tolling classes, based on axles count and height. Used variety of computer vision techniques, from background subtraction to Huff transform. Deployed on 200+ toll lanes and replaced 3x more expensive embedded system.

Led development of a futures spreads charting tool, that allowed portfolio managers and analysts quickly select and pivot various combinations of price timeseries (natural gas contracts):
• Developed vectorized computation library that speeded up spreads and seasonal strips computation tenfold (pandas, numpy)
• Implemented REST backend with load balancing to interact with UI (Flask, gunicorn)
• Led whole application development by iteratively communicating with UI developer and end users, adding/updating requirements, gathering feedback and presenting new functionality
• Deployed “how-to” guides, along with detailed documentation into centralized wiki

Developed deep learning based visual pattern recognition library, which helped to automate trading ideas and report signals on hundreds of instruments, helping traders to focus more on detailed research of revealed patterns.
Development workflow:
1. Timeseries conversion into graphical charts, with manual and semi-manual patterns labeling
2. Testing variety of deep learning approaches
3. Model training and fine-tuning
4. Core library and application development
5. Flask server deployment on AWS EC2 instances

Developed 2 different backtesting applications (in different companies).

Application #1: Strategy testing and optimization framework, with signals generation and performance evaluation on the fly. Framework used Monte-Carlo simulations to pick features' combinations based on the trader-defined universe. These combinations were used to generate series of signals, accounting for patterns' potential biases and create performance report.

Application #2: Core library took holdings and returns series and applied pre-configured metrics to them, generating reports about the strategy, including recommendations based on pre-defined strategy evaluation rules. Library was incorporated into a Flask server, which handled UI requests.

AWS-based Python application with React front-end, that helps track family finances, connects to external bank accounts, aggregates transactions for different reports, categorizes spendings, and produces a future outlook.

Developed and deployed data acquisition framework (pandas, selenium, requests, beautifulSoup) , which included 3 main steps to get publicly available data sets:
- web scraping or REST API download,
- data cleaning: missing data removal, outliers and inconsistencies detection, manually pre-configured checks
- data checks/alerts: helped to mitigate risk of biased/incorrect forecasts
Framework added a structure to companies data acquisition strategy, which allowed to systematically approach to most of the data feeds business had.
Approach was tested in production on more than 50 data sources, which were feeding production models.
Framework also provided unified data access API for analysts to work with in jupyter notebooks.

Integrated MetaTrader4 trading and analysis platform with python Tornado analysis server, which was performing complex patterns search and generating attention signals.
Implementation steps:
1. Research on the integration approaches from MT4 side
2. Socket protocol implementation based on Twisted (python)
3. Load testing strategies scheduling and constant health checks

A risk quantitative framework for weather derivatives pricing uses advanced machine learning, statistics, and probability theory. The framework is written in Python and is used as the main backbone for the company pricing strategy.