Engineered for Deterministic Accuracy & Absolute Privacy
PDFGridly was founded to solve a glaring flaw in modern financial data extraction: generic converters either scramble multi-line table rows or entrust confidential balance sheets to generative AI models that hallucinate figures.
The Core Technical Problem with PDF Tables
A Portable Document Format (PDF) file does not store tabular tables, rows, or columns. It is fundamentally a canvas drawing instruction set containing geometric coordinates (\(X, Y\)) for typography glyphs, vector paths, and raster images.
When accountants or developers convert a statement or invoice into Excel, standard tools rely on one of two flawed paradigms:
Traditional tools read lines sequentially. When an item description wraps onto multiple lines or numbers float between margins, columns shift horizontally and rows corrupt downstream calculations.
AI vision models probabilistically predict numbers. In accounting, a single altered digit or omitted negative sign is an audit disaster. Furthermore, uploading tax documents or bank records to public AI APIs exposes sensitive PII.
The Coordinate-Based Extraction Pipeline
PDFGridly replaces probabilistic guesses with deterministic geometric math.
Vector Token Extraction
Directly extracts digital character glyphs along with their precise two-dimensional physical bounds (\(x, y, w, h\)) without lossy image downscaling.
Baseline Clustering
Clusters words into horizontal baseline rows and maps multi-line descriptions dynamically to their corresponding transaction row anchor.
Column Projection
Determines column boundaries by computing vertical density histograms across headers and values, segregating summary blocks (totals, taxes) from product rows.
Your Financial Documents Never Touch a Hard Drive
Financial statements, payroll pay stubs, and W-2 tax forms contain strictly confidential data: Social Security Numbers, corporate bank balances, employee salaries, and vendor payment terms.
In-Memory Processing: Uploaded PDF streams exist exclusively in volatile RAM buffers allocated to the active Node.js request handler.
Instant Ephemeral Purge: The moment your JSON extraction response or XLSX spreadsheet is streamed to your browser, memory buffers are explicitly unreferenced and recycled by the runtime garbage collector.
No AI Model Training: We never save, log, store, or sell document contents or user telemetry to third parties or model training pipelines.
Authorship & Engineering
PDFGridly is built and maintained by Clara Makes, an independent software engineering studio specializing in high-performance document tools, modern web applications, and privacy-preserving data extraction pipelines.