Building next-gen AI tools for dental professionals

- Built R&D division from ground up - 8 developers and data scientists - working across many of company's key areas to create AI-based decision-making services.- Owning R&D pipeline, envisioning and designing data products, getting buy-ins from CEO and other executives.- Oversight of company's business logic services.- Participation in company’s architecture design.- Advising and coordinating other tech teams and business users.Some of the things we’ve built:- Intelligent pricing model, differentiated across acquisition channels and easily configurable by non-technical users. Pricing combines expert tuning with data-driven AI-based decisions.- Booking and user lifecycle forecasting services, which now became a central part of company's business planning activities.- Company’s big data stack. We receive price data at ~200k events/second, or ~4 TB/day, from backend services and run various predictive models over it. Our stack consists of Kafka + Flink topologies running custom-built streaming ML algorithms, with in-memory Aerospike cluster as a view database.- Visual object recognition system, used for classifying hotel photos.- Matching system, which finds physically same objects in a noisy dataset.- Yield Per Click forecasting model, and derived automated bidding algorithm. It increased marketing efficiency by up to 50%.- Hotel ranking and recommendation models.- Data-driven metasearch bidding strategies.- Company’s split-testing (A/B) system.In most of these projects I was responsible for idea, design and prototype.For smaller services and research we use Python's scientific stack (numpy, scipy, pandas, sklearn, xgboost, TensorFlow, etc). For high-velocity data analytics we use Scala (Spark, Flink).Models are exposed as REST endpoints or run as event-driven actors. Sometimes we also add Go layers. Data is sourced from Kafka streams or relational databases (postgres, vertica).

Worked on Python (django+postgres) and Go backends. We maintained speed and availability in high load setting and were able to scale well.My work primarily focused on designing, implementing and maintaining business logic and decision-making systems.

Created:— pybacktest, an open-source backtesting package in Python and pandas.— Firm's fully automated distributed trading infrastructure. (using Python, SQL, C#, Cython, ZeroMQ).— HFT execution platform with direct access to RTS exchange (now part of Moscow Exchange). Platform was certified with the exchange.— Analytical and visualization tools to be used by other analysts and traders (using mostly Python/matplotlib and R).— Multiple new quantitative trading strategies for RTS index futures and USD/RUR futures.— Family of highly efficient calendar spread strategies for commodity futures markets.— HFT market-marking algorithms for RTS index futures.— Several novel approaches to generating trading strategies based on multistep machine learning pipeline.— Methodology for quantifying and measuring amount of data snooping that happened during trading strategy development.— New metrics and visualisation techniques to measure amount of alpha contained in a signal.