Automate Rent Roll Reconciliation with Custom AI

Most asset management teams start by trying generic OCR tools. A tool like Adobe Acrobat Pro can extract text from a PDF, but it cannot understand the structure of a rent roll. It fails to distinguish between 'Base Rent' and 'CAM Charges' if the column headers are inconsistent across documents, leading to incorrect data extraction.

Consider an asset manager for a 250-property fund. Each month they receive rent rolls in different formats: clean Yardi exports, scanned PDFs with handwritten notes, and Excel files with merged cells. An analyst spends the first week of every month manually keying this data into a master spreadsheet. This process introduces an average error rate of 3% and delays reporting by a week.

Template-based data extraction tools like DocuParser fail because of this variability. They require a fixed template for each document layout. When a property manager adds a new column for 'COVID-19 Relief' or changes a date format, the template breaks. The analyst is forced back into manual entry, which defeats the purpose of the tool.

Syntora would approach the problem by first conducting a discovery phase to understand your specific document types and reconciliation rules. This would involve collecting a representative sample of 20-30 rent roll documents, covering all major formats you receive. We would use Python's pdfplumber library to extract raw text and table data, analyzing the structural variants present across your documents. From this, we would define a precise mapping from source document fields to the target fields required by your accounting system, such as Yardi or MRI.

The core logic for intelligent document understanding would utilize the Claude API. Syntora has experience building similar document processing pipelines using the Claude API for financial documents, and the same pattern applies to rent roll documents. We would feed the extracted text to Claude with a carefully engineered prompt to identify critical entities like tenant names, lease IDs, rent amounts, and service charges. Using a structured output model, the API would return JSON data, which we would validate with Pydantic to ensure data integrity. This approach is designed to handle variations in document format because Claude processes context and meaning, not just an entry's fixed position on a page.

The validated data would be written to a Supabase PostgreSQL database, providing a reliable and scalable data store. We would build a FastAPI endpoint to manage the ingestion of new rent rolls, designed for secure and efficient uploads. The system would be engineered for rapid processing, aiming to provide structured data from a PDF upload within a few minutes per document. This service would be deployed on AWS Lambda for cost-effective, scalable operation, with typical hosting costs expected to be modest for processing hundreds of documents monthly.

As part of the engagement, Syntora would develop a simple front-end application, potentially deployed on Vercel, to facilitate a 'human-in-the-loop' review process. This interface would allow your team to easily review the extracted data alongside the original PDF. It would highlight any identified discrepancies against your accounting system's existing data, for example, a rent amount difference exceeding a predefined threshold. This step ensures that data can be quickly verified and corrected before final import, maintaining high accuracy and control over your financial data.