Andrii B.

Project overview:

A mission-critical regulatory compliance system for automated EU sanctions screening of maritime vessel voyages at Metinvest Group — a $12B+ enterprise with global shipping operations. The system replaced a manual, error-prone compliance review process with a fully automated, auditable pipeline capable of processing hundreds of voyages annually. Regulatory defensibility was a hard requirement: every automated decision had to be traceable, explainable, and audit-ready.

Responsibilities:

• Designed the complete pipeline architecture: SAP (IMO trigger) → Azure Blob Storage → PDF ingestion → OCR extraction → rule-based EU sanctions screening → LLM scoring for ambiguous cases → auto-decision or Teams escalation (human-in-the-loop) → writeback to SAP.
• Implemented a two-layer compliance engine combining deterministic rule-based logic for clear cases with LLM-based probabilistic scoring for ambiguous vessel records.
• Built a full audit trail subsystem ensuring every decision — automated or escalated — was logged with decision rationale for regulatory review.
• Integrated with Azure Blob Storage (Managed Identity), SAP, and Microsoft Teams API for human-in-the-loop escalation workflows.
• Engineered the system for zero third-party data exposure — all inference ran on self-hosted infrastructure within the corporate perimeter.

Achievements:

• Automated screening of hundreds of vessel voyages per year — compliance team shifted from full document review to exception handling only. • Two-layer architecture (rule-based screening + LLM for ambiguous cases) reduced false positives while maintaining zero-miss accuracy on sanctioned entities. • Full audit trail on every decision ensured regulatory defensibility — zero findings during external compliance audits. • End-to-end pipeline from SAP trigger to auto-decision writeback achieved without any manual handoffs for clear-cut cases.

Technology stack:

Project overview:

A large-scale agentic document processing system for Metinvest Digital, designed to ingest, recognize, and process accounting and transportation documentation from contractors across Europe — Italy, Portugal, and Poland. The system handled hundreds of thousands of documents in real-time, integrating OCR transformation with downstream SAP financial workflows. A significant legal constraint was navigating restrictions on digital signatures across jurisdictions, requiring a purpose-built compliance workaround.

Responsibilities:

• Developed the complete document ingestion and processing pipeline from scratch: multi-source document intake → OCR extraction → structured data transformation → SAP financial system integration.
• Designed the agentic orchestration layer for parallel processing of high document volumes across multiple source countries and document formats.
• Engineered a compliant workaround for cross-jurisdictional digital signature restrictions involving duplicate certificate issuance
• Integrated with Azure Blob Storage and OpenText for document storage and lifecycle management.
• Built real-time processing pipelines capable of handling peak document loads without throughput degradation.

Achievements:

• Built from scratch within the internal R&D department — zero prior infrastructure to build on. • Processed hundreds of thousands of documents in real-time with full SAP integration for downstream accounting workflows. • Engineered a legal workaround for digital signature restrictions across EU jurisdictions using duplicate digital certificate issuance — passed regulatory review. • System subsequently adopted as the standard document processing infrastructure for European contractor operations.

Technology stack:

Project overview:

An ultra-fast, privacy-first multi-agent backend for a high-load conversational AI platform similar to Character AI. The system required simultaneous compliance with GDPR and the EU AI Act, demanding that data sovereignty, age verification, and content moderation be architectural first-class concerns — not afterthoughts. The core challenge was achieving sub-millisecond agent response times without relying on heavy frameworks that sacrifice observability and control.

Responsibilities:

• Served as Fractional CTO — defined the full technical architecture, made all stack decisions, and led end-to-end backend delivery.
• Designed and implemented a high-performance multi-agent orchestration core optimized for concurrent, low-latency roleplay sessions.
• Built custom memory and context management logic with contradiction detection to maintain character consistency across sessions.
• Integrated identity and age verification providers as a core compliance pipeline, not a post-release patch.
• Architected the system around Privacy-by-Design principles to satisfy GDPR and EU AI Act requirements without performance trade-offs.

Achievements:

• Delivered a fully custom Python backend from scratch, without heavy frameworks, achieving maximum throughput and observability at production scale. • Implemented Privacy-by-Design architecture ensuring the client retains full status as data owner — zero third-party data exposure by design. • Built a custom multi-agent memory layer with Contradiction Resolution logic, enabling consistent, coherent character personas across long sessions. • Achieved full GDPR and EU AI Act compliance from day one, including integrated identity and age verification pipelines with audit trail support.

Technology stack: