From reading text to understanding a page.

Contour detection locates the page. Affine homography transformation corrects perspective and skew. CLAHE — Contrast Limited Adaptive Histogram Equalization — normalises contrast locally rather than globally, which is what recovers a faded carbon copy photographed under a single tube light without blowing out the parts that were already legible. This stage runs before anything is read, and it is the largest single accuracy lever in the pipeline.

The visual hierarchy of the page is identified — tables, header blocks, stamp seals, signature fields, key–value pairs and line-item grids — and a reading order is established across them. This runs before recognition, because knowing a block is a table column changes how its contents should be read. Flatten the page first and the structure that made it meaningful is already gone.

The recognition core. A vision-language transformer fine-tuned on Devanagari matras, half-characters, complex joint ligatures and mixed-code Hinglish — the constructions where engines built for Latin script produce character truncation and disordered output. Text is recovered together with its position, not as a flat stream stripped of geometry.

Extracted blocks are normalised into typed JSON schemas — tax invoice, land record, KYC form — and vector embeddings are generated alongside them, so the same pass that digitises a document also makes it available to semantic search, automated accounting and downstream LLM analysis without a second extraction.