How AxLiner's Built
A fine-tuned document engine for handwritten tables.
AxLiner uses a 7B-class vision-language model in the Qwen2-VL and olmOCR family, then wraps it with table recovery, batch processing, durable metadata, and Excel export logic.
Model stack
Vision encoder
Pages are normalized into visual tokens so ruled tables, handwriting, stamps, shadows, and phone captures stay readable before extraction.
Instruction tuning
The OCR path is shaped around Qwen2-VL and olmOCR-style document prompts, then tuned for handwritten text, table boundaries, and spreadsheet output.
Schema alignment
Rows, headers, merged cells, totals, and column relationships are rebuilt before export so the workbook behaves like a table, not loose text.
Batch orchestration
Redis-backed queues, durable file metadata, and retry-aware workers keep multi-file conversion recoverable when users upload real batches.
Compact operator-style signals for the conversion path.
Handwritten specialist
The extraction prompt and review path are tuned for handwritten tables, messy notes, invoices, forms, and scanned PDF pages.
Batch-first product layer
Jobs, files, downloads, and review states are kept together so a user can process many pages without losing track of outputs.
Spreadsheet review loop
The output is built for correction, comparison, and final download rather than a one-shot OCR text dump.