InvoiceAI: Automated OCR & Warehouse Accounting Intelligence

The Challenge: Manual data entry for warehouse shipments is a major bottleneck, costing businesses ~10 minutes per invoice and hundreds of thousands in administrative overhead. Fragmented vendor nomenclature and handwritten corrections make standard OCR solutions fail, leading to inventory discrepancies.

The Solution: I engineered an end-to-end AI pipeline that transforms physical invoices into structured data ready for warehouse management systems.

Multimodal OCR Engine: Implemented a hybrid system using Gemini 3 Flash and Mistral OCR with a Google Cloud Vision fallback. It handles handwriting, crossed-out items, and manual corrections with 99% precision.

Intelligent Entity Matching: Built a sophisticated matching layer using Fuzzy Search and Vector Embeddings to correlate non-standard supplier names with internal nomenclature.

Automatic Document Grouping: AI-driven logic that automatically assembles multi-page scans and photos into coherent single-invoice objects.

Seamless ERP Integration: Developed a robust interaction layer with Syrve (iiko) and 1C (OData) for automated document creation and real-time product/supplier synchronization.

Human-in-the-Loop Interface: An administrative dashboard for manual verification of high-risk items, featuring inline editing and synonym management.

Telegram-Based Data Intake: A specialized bot for field staff to upload photos directly into the processing queue.

Operational Efficiency: Reduced invoice processing time from 10 minutes to <30 seconds.

Cost Savings: Projected annual savings of $9,000+ per accountant by eliminating manual entry.

Accuracy: Eliminated human errors in price, VAT, and unit conversion calculations.