How TRBank, Inc. (A Rural Bank) transformed paper loan records into structured credit intelligence
5,100 loan records. 194 fields per loan. Automated validation across the full corpus.
TRBank, Inc. (A Rural Bank) is a tech-forward rural bank serving MSMEs across the Philippines. Founded in 1956 and backed by Asialink Finance Corporation and Stem Financial, it lends through a network of 9 full branches, 11 Branch Lite Units, and a Head Office.
The bank set out with a clear goal: to make four years of its own lending history usable. Between 2022 and 2025 it had approved thousands of loans, yet every one was recorded in its own way, with no common format and no shared structure to bind the archive together.
A credit accommodation record at TRBank is the source-of-truth file for every approved loan. The archive runs to thousands of records accumulated through years of lending operations.
Each record brings together the loan application, underwriting analysis, collateral appraisal, bank-statement review, scorecards, bureau report, typed forms, handwritten notes, scanned documents, and photographs of originals.
One folder often runs to several hundred pages. The full archive runs to thousands of folders, organised by month and year of release. The shape will look familiar to anyone who has worked credit at a legacy lender.
Before the project, answering even basic questions about historical lending activity required manual retrieval and review. The operational bottleneck was not the absence of data; it was that the data was locked in physical and scanned records that could not be searched, compared, or reused at scale.
TRBank selected Kita after testing confirmed that high-accuracy extraction was achievable on the Bank’s own lending documents, field taxonomy, and scanned historical records.
The pilot showed that Kita could process TRBank’s source images against the Bank’s 194-field schema and produce structured outputs with strong field-level accuracy. This gave the team confidence to move from testing to a full archive run.
Kita also stood out for practical reasons that mattered to TRBank: focus on the Philippine market and familiarity with banking documents, local IDs, peso-denominated records, and rural bank operating realities.
Speed became the decisive proof point. The archive processing was completed in less than a week, with the core production run taking no more than three days. With Kita’s capabilities, the results showed that once source files, schema, and validation rules are ready, historical archive conversion can happen on timelines that change what is operationally possible.
Tagalog handwriting, Philippine government IDs, peso amounts, local bureau formats, all handled out of the box.
TRBank’s own field schema preserved verbatim. Seven scorecard template families. 26,568 exception categories.
Faded scans, photographs of originals, and partial crops were handled through a validation-oriented workflow designed to reduce guesswork and flag ambiguity for review.
The project converted historical lending records into structured outputs that can support faster retrieval, better portfolio visibility, and more informed operational decision-making.
The pipeline transformed each TRBank loan record into a structured 194-field output. The schema spans 19 thematic sections covering loan application metadata, borrower identity, residential and employment history, financial position, source of income, collateral appraisals, bureau checks, bank-statement summaries, and scorecard outputs. Field names were preserved from the TRBank specification.
Alongside the main pipeline, two specialized extractors handle the highest-value surfaces of a credit folder.
Credit folder
Typed forms, handwritten notes, scanned documents, photographs of originals. Several hundred pages per folder.
Main pipeline
194 fields across 19 sections per record.
Scorecard extractor
Classifies each scorecard page into the right TRBank template; pulls the numeric scoring grid.
Remark extractor
Captures handwritten margin notes and Remarks-box entries; classifies into 26,568 exception categories.
Structured record
Flat-file deliverable. 194 fields, scorecard grids, remark categories, bank-statement snapshots.
Four years of TRBank's lending archive, in flat-file form, ready for direct use.
Numbers below describe the delivered corpus.
The project matters because it turns historical lending information into operational intelligence that can support faster, simpler, and more informed lending experiences.
Most legacy lenders are sitting on a similar archive. Decades of bound credit folders. Tens of thousands of paper applications. Photocopies, photographs, scans of originals from before the digital underwriting era. It has been carried as overhead: storage, retrieval, compliance hold. The signal value of what it contains has been written off as unqueryable.
Once a lending archive is structured and validated at the field level, historical credit decisions become easier to retrieve, compare, and analyze. Portfolio analytics can be grounded in actual historical records. Policy review can move from anecdote to evidence. The Bank’s own lending experience becomes a reusable input for future credit decisioning.
For traditional lenders, this kind of archive is especially valuable. The lending book is the moat: the relationships, the local credit signal, the decades of what actually defaulted versus what didn’t. Once it is structured, the archive stops being inventory in a filing room and becomes the institution’s largest underwriting asset.
TRBank showed that historical credit files are not just operational records. Structured correctly, they become a competitive advantage.
“TRBank’s lending archive was, in practical terms, unqueryable: thousands of credit folders accumulated over years, each running to several hundred pages of typed forms, handwritten notes, collateral appraisals and bureau reports.
Working with Kita, a specialist document AI platform built for lenders, we processed 5,100 credit folders spanning 2022 to 2025, structuring over 450,000 data points with independently verified accuracy above 97%. PHP 8.77 billion of approved loan value is now fully queryable for the first time: the bank’s own ground-truth record of what it lends, to whom, on what terms and against what collateral.”
Curious what this could look like for your bank?
Speak with our team.
Or write to hello@kita.ai