Adam Ivansky

I architected and engineered an end-to-end real-time data integration between a white-label ticketing platform and Snowflake. The core architecture features a high-throughput, serverless ingestion pipeline using AWS API Gateway and AWS Lambda to process live webhook events, streaming them directly into Snowflake via Snowpipe. To complement the real-time stream, a custom API scraping solution was implemented to ingest static product catalog data. Inside Snowflake, data normalization and flattening were automated using Dynamic Tables, delivering low-latency, high-performance data models. This robust data foundation directly powers critical operations dashboards used for real-time seating inventory management and dynamic ticket pricing.

I spearheaded the design and implementation of an enterprise-grade agentic AI platform leveraging a self-hosted Mixtral Mixture of Experts (MoE) LLM to automate advanced lead intelligence. The system utilizes an event-driven function-calling framework to execute semantic searches across Snowflake data warehouses and distributed APIs. By orchestrating a secure retrieval-augmented generation (RAG) pipeline, the autonomous agent ingests multi-source data payloads and performs contextual text synthesis to generate structured, intent-driven conversation starters. This private LLM deployment mitigates data leakage while serving low-latency, production-ready GenAI insights directly into Salesforce, optimizing the outbound sales funnel through automated, cognitive pre-call research.

I developed and managed an enterprise-grade web platform on AWS ECS that served Streamlit business performance dashboards backed by robust Snowflake data models. To ensure strict security compliance, the architecture integrated Okta and Azure AD Entra for federated authentication and fine-grained permissions management. Deployed via a modern GitOps framework, the platform empowered developers to contribute via a unified Python GitHub repository, utilizing GitHub Actions pipelines to orchestrate automated deployments across fully isolated development, UAT, and production environments.

The aim of the project was to increase the click-through rate of eBay coupon campaigns via the use of machine learning. The development of the algorithm was successful, and it was measured to generate a 20% lift in click-through rate and IGMB.

The early version of the algorithm was based on the multi-armed bandit. Later versions made use of contextual NLP-based multi-armed bandit. The algorithm was developed using a combination of Teradata SQL and Python. I also developed an interactive Tableau dashboard in order to monitor the function of the algorithm and to measure the KPI lift that the algorithm was bringing.

The project aimed to develop a machine vision algorithm capable of detecting pedestrians in front of a vehicle by analyzing the input from the vehicle camera. The algorithm is now fully functional and embedded into several newer vehicle models, including the GM 2019 Chevy.

The machine learning algorithm we decided to use was the AdaBoost cascade classifier combined with a deep neural network. We wrote the training application from scratch in C++. Training had to be multithreaded in order to be efficient. Testing and validation were done in Python. A large database of annotated video data was used for algorithm training.

Precise prediction of the total final loss after a client's default is key to reducing the risk associated with different loan products.

I developed a model that relied on the loan-to-value ratio and the value of the collateral. It was done using a combination of SAS and Microsoft SQL Server. The development of the model required extensive data cleaning and data quality testing.

Involved in the development of a recommendation engine based on a collaborative filtering model. The engine was capable of recommending even the products that a given customer did not necessarily buy in the past. The solution was implemented in PySpark and was based on MLlib, Spark's machine learning (ML) library.

Developed an ETL in PySpark to transfer data from Amazon Redshift into an Amazon S3 data lake. I also developed code for customer-level data aggregation and historicization. Finally, I assessed data quality and investigated and remediated data quality issues.