Viacheslav V.

Project overview:

Designed and delivered from scratch a large-scale FinTech module called AIR (200k+ LoC) - a Cash Application system within an accounts-receivable platform originally built as a Ukrainian with successfull following acquisition. The system integrates directly with ERPs to automate processes The AI component handles the matching of incoming payments with accounting documents and invoices. It processes documents from diverse sources - email body, attachments, scanned receipts, bank files, and direct ERP feeds. End-to-end document categorization, OCR recognition, named entity extraction, and probabilistic matching of documents to specific payments or customers. The module contributed to ARR growth from zero to millions.

Responsibilities:

• Conducted discovery and architecture design on how to utilize AI/LLM to improve the matching of incoming funds with business documents.
• Developed a full ecosystem for reading, recognizing, and targeting invoices, including calculating matching probabilities using OCR, NER, and LLM pipelines.
• Managed the integration of multiple OCR engines and Large Language Models to parse text, tables, handwritten receipts, and document numbers - selecting optimal models per document type for accuracy and cost.
• Implemented the core business logic (in Java) that processes LLM outputs to close payments and update the ERP in an event-driven, Kafka-centric architecture.
• Owned reliability and evolution of services in a containerized Kubernetes environment with end-to-end production operations.

Achievements:

• Achieved an overall matching rate of 80-90%, with AI specifically responsible for 60-70% of those successful matches. • Improved performance of critical processing functionality by ~5-6x, reducing support tickets ~2-3x and improving customer satisfaction. • Successfully maintained PCI DSS compliance by hosting sensitive payment data and OCR processes on bare metal/on-premise servers. • Executed large-scale refactors and systematic AI-assisted test generation, achieving 95%+ test coverage. • Reduced needed manual labor by ~85%:

Technology stack: