Architecture & Pipeline Specifications

Bank Statement Parser is architected around a deterministic, four-stage extraction and validation pipeline. Every stage executes in-process with zero outbound telemetry, zero heap allocation in the critical path, and strict mathematical balance verification before emitting ISO 20022 messages.

Four-Stage Processing Pipeline

Deterministic Pipeline Execution Trace

The trace below demonstrates the sub-millisecond execution lifecycle of the four-stage engine processing a multi-page PDF statement into validated ISO 20022 CAMT.053 XML:

Stage 1: Zero-Copy Stream Ingestion

The ingestion subsystem interfaces directly with the host filesystem via memory-mapped files (mmap), bypassing kernel-to-userspace buffer duplication. Byte slices are analyzed using non-backtracking magic-byte sniffers:

• PDF Documents: Validates file header prefix %PDF- and reads the cross-reference table (xref) directly from the trailer offset.
• OFX Streams: Sniffs OFXHEADER SGML tags or XML prologue <?xml version="1.0" to branch into streaming parser trees.
• SWIFT MT940 / MT942: Checks for field delimiter sequences (:20:, :25:, :28C:) to route into deterministic line tokenizers.
• CSV Delimited Files: Evaluates entropy across candidate delimiters (comma, semicolon, tab, pipe) over the first 50 rows to infer column geometry without full-file scans.

Stage 2: Spatial Layout Tokenization

For unstructured formats such as digital PDF statements, standard text extraction destroys critical vertical alignment. Bank Statement Parser utilizes a deterministic coordinate-clustering engine:

• Glyph Extraction: Maps character matrices, font bounding boxes, and baseline coordinates from content stream operators (BT, ET, Tj, TJ, Tm).
• Row Clustering: Groups glyphs with overlapping vertical tolerances into discrete lines, preventing footnote and header collision.
• Column Geometric Alignment: Computes vertical whitespace gutters to delineate Booking Date, Value Date, Narrative, Amount, and Running Balance columns.
• Multi-line Narrative Stitching: Heuristically associates wrapped description text back to the initiating transaction row.

High-assurance statement parsing: zero-copy memory mapping, deterministic tokenization, and ISO 20022 message synthesis.

Stage 3: Field Normalization & Standard Mapping

Extracted raw string tokens are converted into strongly-typed canonical primitives:

• Date Canonicalization: Resolves 24 regional date formats (including DD/MM/YYYY, MM/DD/YYYY, YYYY-MM-DD, and multi-lingual month names) into ISO 8601 UTC timestamps.
• Monetary Amounts: Parses European period/comma inversion (e.g., 1.234,56 vs 1,234.56), preserving scale as exact 64-bit fixed-point integers to eliminate IEEE 754 floating-point rounding drift.
• Account & Bank Identifiers: Verifies ISO 13616 International Bank Account Numbers (IBAN) using mod-97 arithmetic (ISO 7064) and validates ISO 9362 Business Identifier Codes (BIC).
• Transaction Code Classification: Maps proprietary transaction labels to standardized ISO 20022 Bank Transaction Codes (BTC), distinguishing card settlements, direct debits, credit transfers, and fee deductions.

Stage 4: Mathematical Validation & Schema Synthesis

Before any extracted statement is emitted, the engine evaluates strict balance invariants to guarantee data integrity:

Upon invariant satisfaction, the engine streams the normalized data tree into target schemas:

• ISO 20022 CAMT.053: Generates conforming camt.053.001.08 Bank-to-Customer Statement XML messages ready for direct core banking and treasury ingest.
• Canonical JSON: Emits clean, strongly-typed JSON with ISO 8601 dates and exact decimal numbers for downstream API consumption.
• Reconciliation CSV: Serializes standardized transaction rows with consistent column headers for enterprise ERP systems.