Solutions Architect and Developer in Montreal, QC, Canada

TASK
Improve the web search experience for Premiumbeat, a stock music provider, to surface matching available content when customers search for well-known popular names.

SOLUTION
Developed an ML-based analytic system, processing daily search stats and using Spotify API and Music Brainz databases to tag offered tracks matching popular artists and songs. As a result, searching for "zeppelin" returns a few dozen classic rock tracks, matching them by the desired style rather than by name.

TASK
Enable web search results to be reordered based on customers' geo regions, current and seasonal trends, and collaborative filtering.

SOLUTION
I led a solution to implement a data analysis pipeline feeding a collaboration filtering ML model to achieve better surfacing of the video content. This improved the first-line click-through rate by 15% and the first-page click-through rate by 30%.

TASK
Create a system capable of answering virtually any fact-check type question and complement existing embedded car assistants with this capability. Almost single-handedly, I designed, developed, and deployed a POC for such a system. It used a unique process to search for public knowledge and extract an answer to a given question by using an advanced AI model.

KEY ACTIVITIES
• Designed a POC solution.
• Designed and coded two microservice APIs, using Node.js and Python running on Azure Functions (serverless).
• Integrated with Google and Bing APIs.
• Developed an API server to host a PyTorch DNN model.
• Deployed all components into Azure Cloud, using Web Apps, Functions, VM, API Management, and Insights.
• Developed and deployed a BDD test suite.
• Set up a statistics data stream to a Power BI dashboard.

TASK
Replace the business backbone database holding two billion+ assets with no downtime for the company.

SOLUTION
Designed and led a no-downtime production migration from MySQL to a new Cassandra-based media database with the following criteria:
• Two billion assets, scalable to 10 billion.
• Candidate selection through POC.
• Peak performance of 120,000 TPS.
• Geo-distributed.
• No migration downtime.

TASK
Conduct R&D research for a system to detect videos shot in lower resolution and later upscaled by software.

SOLUTION
Conducted research and built a POC service capable of performing this task.

KEY ACTIVITIES
• Ran paper research and a series of experiments.
• Transformed scientific research findings into real-world algorithms.
• Trained and applied a variety of ML models.
• Demoed the solution and results to the executive team.

TASK
Create a tool that can detect musical instruments in a given soundtrack.

SOLUTION
Researched the state of the art and designed a new system using a VGGish model trained on a YouTube-8M dataset. The model has been used as the first stage for transfer learning where a new classification model has been built to classify instruments.

CHALLENGE
An end-user companion mobile app displaying a car's telemetry should have around 200 ms overall latency for the whole trip: from hardware to IoT cloud to storage to API to mobile app. The traffic comprises one million cars reporting every five seconds, making it 200,000 per second.

SOLUTION
Designed and led the development of Azure-based serverless cloud architecture, maximally using power or native cloud streaming components. It provides performance at a great scale with the required reliability for the automotive domain.

Engaged as a lead technical architect, leading the design and MVP estimation for a blockchain-based crypto investment platform allowing social content makers to capitalize on their work and for their subscribers to participate in financing new projects.

Chosen Technology: Blockchain for stake participation and dividends distribution