Luboš Charčenko

One company I was cooperating with, let's call it Xteam, was selling services from various portals.

Once a customer bought a specific service, Xteam went and created an account for the user within the portal, bought the service, and provided it back to the user with a margin, of course.

However, at some point, Xteam wanted to empower clients by enabling users to interact with a portal without handing over credentials for some services.

Lastly, some portals had just one login for all the customers. I invented and built a proxy with multiple functions to solve these issues.

Proxy Functions:
• It allowed the opening of nearly any page through an iframe by adjusting the headers.
• It could perform hidden automated actions, like login without showing credentials to the end user.
• It rerouted all the links back to itself with mapping to the original links, so the proxy was aware of all the traffic from the site opened through an iframe. That way, it could restrict actions within portals and hide or change some parts of the portal, like filtering only cases related to a specific user for single account portals.
• It had logging interactions that could propagate changes back to the main system.

I designed and developed a version of a flight-combining algorithm based on PostgreSQL triggers.

After a new flight is added or an existing one is updated, the database automatically checks the new combination's prices and validity and generates new ones or deletes the invalid ones if needed.

At the time, the database was processing 2,000 updates per second, which cascaded into tens of thousands of updates every second on top of 300 million existing combinations.

For the search, I used cascaded replication across multiple search-optimized PostgreSQL nodes.

I developed a portfolio management system using CNN and RL. The system handles over 1 million updates per second with stateful workers grouping data for aggregated representations, which are then stored in ScyllaDB with over 10 billion rows and growing.

Tech: Python, Kafka, ScyllaDB, Celery, Redis, Kubernetes, Datadog

Data is crunched by a pipeline to create 4D representations of vital features in space to create historical and real-time datasets for convolutional neural networks and deep reinforcement learning.

Tech: Python, Pandas, NumPy

I tried a lot of different architectures, with the best based on CSPDarknet53 ideas. First, I implemented deep reinforcement learning with parallel independent agents training the same network able to handle the whole portfolio at once and distribute capital optimally between investing and hedging.

A secure trader works with predictions and communicates with a broker API; it distributes capital between multiple algorithms and works with multiple trading accounts—monitoring, alerts, and an emergency kill switch with SMS.

Tech: EC2, IB API, AWS 2-way SMS, PagerDuty, Datadog, Python, Asyncio, Pandas, NumPy, Matplotlib, TensorBoard, TensorFlow, Keras, Jupyter

I've written about 200 functional scraping modules using various technologies.

Sometimes websites fight back with CAPTCHA programs, IP, and even recording mouse movements or exact click locations on elements.

Fighting those protections while keeping modules performant and saving as much as possible on traffic forcefully pushes my creativity to an entirely different level.

Technology Stack: Under NDA

I invented a type of software enabling complex web interactions without needing any coding to offload work from the engineering team.

The system uses precoded elements, which extract data, interact with the page, and drag-n-dropped into the site automation plan. The system is designed to bypass CAPTCHA programs or other anti-scraping techniques.

Technology Stack: Under NDA