Roman T.

Project overview:

The project involved designing and productionizing an AI-assisted governance and code review system for a privacy-preserving Data Clean Room platform. The platform was built to support secure data collaboration across isolated client environments, where SQL, PySpark, Scala, and pipeline definitions had to comply with strict privacy, PII, data-contract, and governance rules. The AI assistant was designed to analyze engineering changes before execution, detect potential policy violations, explain risks, and provide actionable recommendations to engineers. Instead of using an unconstrained chatbot approach, the solution combined LLM reasoning with structured platform context, explicit governance rules, deterministic validation checks, and human-in-the-loop approval flows.

Responsibilities:

• Designed the AI assistant architecture for governance-aware review of SQL, PySpark, Scala, and data pipeline changes.
• Integrated LLM capabilities through AWS Bedrock while keeping platform-specific context, policies, and execution boundaries under control.
• Defined validation flows for PII exposure, data contract violations, unsafe transformations, and non-compliant data access patterns.
• Created structured context packages containing platform rules, architecture constraints, repository conventions, and review criteria.
• Implemented a human-in-the-loop review model to ensure that AI recommendations remained explainable, auditable, and safe for production use.
• Designed the productionization path, including backend deployment, IAM boundaries, async validation flows, and operational monitoring considerations.
• Evaluated MCP/MaaS-style integration patterns to expose platform capabilities to AI agents without coupling the solution to a single assistant UI.

Achievements:

The solution was moved from PoC to production-ready implementation in approximately 1.5 months. It helped reduce manual first-line review effort, improved consistency of governance checks, and created a scalable foundation for AI-assisted engineering workflows within a data-sensitive platform. The implementation demonstrated that LLM-based assistants can be safely applied to enterprise data engineering when constrained by bounded context, platform rules, and deterministic validation layers.

Technology stack:

Project overview:

The project involved designing and building a cloud-agnostic buy-side advertising platform (Demand-Side Platform). The platform included high-load real-time bidding components, campaign and targeting management, event ingestion, data processing pipelines, and a reporting layer based on ClickHouse. The system was designed for programmatic advertising workloads where latency, throughput, cost efficiency, and data reliability are critical. The architecture supported real-time bid request processing, campaign targeting, audience and frequency-capping checks, event logging, aggregation, and business reporting. The infrastructure was built to run in a Kubernetes-based environment with Terraform-based provisioning, allowing deployment across different cloud providers or private infrastructure.

Responsibilities:

• Designed the end-to-end architecture of a buy-side advertising platform, including bidder, campaign management, targeting, event ingestion, aggregation, and reporting components.
• Architected the real-time bid stream processing flow with early traffic filtering, geo/IP enrichment, targeting evaluation, audience checks, frequency capping, and budget controls.
• Designed high-load backend services optimized for low latency, predictable memory usage, and efficient request processing under programmatic advertising traffic.
• Built the data platform architecture using ClickHouse for aggregation and reporting workloads, separating operational data flows from analytical queries.
• Designed event logging and ingestion flows for bids, impressions, clicks, conversions, and campaign performance metrics.
• Defined cloud-agnostic infrastructure patterns using Kubernetes, Terraform, Docker, and environment-based configuration management.
• Supported production-readiness activities, including performance tuning, scalability planning, deployment strategy, monitoring, and troubleshooting.
• Worked with product and engineering teams to translate business requirements into scalable technical solutions for campaign delivery and reporting.

Achievements:

Delivered a production-grade high-load advertising platform architecture capable of handling large-scale bid traffic and data-intensive reporting workloads. The platform combined low-latency request processing with a cost-efficient analytical backend based on ClickHouse. The cloud-agnostic Kubernetes and Terraform setup reduced infrastructure lock-in and made the platform adaptable to different client environments and deployment constraints.

Technology stack: