Market intelligence and Credit View work from documents rather than spreadsheets. Both run on the Document Extractor, a pipeline EDEN built for the documents its clients actually have: Indian loan bundles and company filings, often scanned or photographed, in mixed formats and sometimes mixed languages.
Off-the-shelf OCR returns text. Teams need answers: the sanction amount, the EMI, the revenue for the quarter, with the page it came from and a sense of how sure the reading is. The Document Extractor is built to return exactly that.
Why EDEN built its own
- Loan files and filings are keyword-dense: PAN, IFSC, chassis number, LTV, EMI, EBITDA. A retrieval method tuned for exact terms finds the right passage more reliably than a general-purpose search over these documents.
- Indian number formats — 1,00,000, 48.60 lacs, Rs. 12.5 Cr — and date styles have to be understood, not just read.
- Every value needs a source. The pipeline keeps the document, the page, the original text and a confidence with each value, so the output can be checked and audited.
- The knowledge bases are EDEN's own and grow with clients: a parameter or a measure is added as data, and the extractor, the checks and the exports pick it up.
How it reads a document
- Read every page. Digital PDFs and spreadsheets give their text directly. Scans, photos and image-only pages are read by a vision-capable AI model used as the OCR engine, so a photographed KYC document or a scanned invoice takes part in the same pipeline as a digital page. Credit View reports how many pages were read each way.
- Classify the page. In Credit View, every page is labelled with one of 46 document types — sanction letter, application form, CAM, KYC documents, proforma and tax invoices, insurance policy, RTO forms, NACH mandate, bank statement, bureau reports, field and tele-verification reports and more — by scoring its text against the knowledge base. The labels drive the document checklist and tell the extractor which pages deserve a vision pass.
- Index for retrieval. Pages are split into bounded chunks and indexed with BM25 keyword ranking, so for any question the most relevant passages across the whole bundle are found in milliseconds, with the page recorded for each.
- Extract against a knowledge base. For each group of parameters, a focused query retrieves the best passages and an AI model extracts values into a fixed schema, group by group rather than page by page. Credit View works from a knowledge base of up to 250 loan parameters in 17 groups; Market intelligence from 60 financial measures in 12 groups, each with its aliases, unit and valid range.
- Normalise every value. Money, percentages, counts, dates, identifiers, names and addresses are each parsed by type: Indian grouping and lakh/crore words become numbers, dates become real dates, identifiers are tidied for comparison. Market intelligence also puts figures on common units and fiscal periods, derives margins and ratios where only the inputs are reported, and converts currency with a published ECB-sourced rate whose date is shown.
- Capture the tables. Some facts live only in tables. Credit View extracts the repayment schedule, bank-statement transactions, bureau account history and existing loans as tables, from digital and scanned pages alike, because the EMI arithmetic, the cash-flow cover and concurrent exposure depend on them.
- Check the result. Credit View compares each parameter across documents with rules suited to its type, lists mismatches by severity, checks the bundle against the required-document list and evaluates twelve lending-policy rules. Market intelligence validates each measure against its range and flags it against the benchmark range and the peer group.
- Keep the source. Each value carries its document, page, page type, the original text and an extraction confidence. The results pages and the workbooks show them, so a reviewer opens the page rather than trusting the number.
What it means for your team
- A forty-document disbursal file is one upload, not a morning of keying. The work that remains is reviewing what the extractor found and flagged.
- Disagreements between documents surface before disbursal, with both pages in front of the reviewer.
- Scanned and photographed pages are first-class input, not an exception that goes back to manual processing.
- Case data leaves in the shape your systems expect: a seven-sheet audit workbook and a CPA mapping of 174 columns matched to the client's credit processing template.
- Processing effort is a choice: three tiers trade cost for depth on difficult bundles, and the cost of every run is estimated before and shown after.
- Filings from up to ten companies become one comparison on consistent measures, with the outliers already marked.
Processing tiers
- Lite
- Light bundles, lowest cost. Best for simple cases: clean PDFs, few scanned pages.
- Medium
- Balanced, recommended for most cases. Good accuracy on standard commercial-vehicle, equipment and home-loan bundles.
- High
- Complex bundles, highest accuracy. Best for handwritten KYC, dense CAM memos, ambiguous scans and mixed-language documents.
What this does not do
- Low-quality scans, handwriting and unusual layouts reduce accuracy. Check the confidence shown with each value.
- AI models can misread a value. Treat extracted figures as a first draft for review, not a final record.
- Market intelligence keeps the figure, unit, period and confidence, but not the page it came from.
- How uploaded documents are handled is described in the Privacy Policy.