Carmel Limcaoco keeps a folder on her phone filled with the kinds of documents that make legacy software systems sputter and quit. A bank statement with ballpoint annotations scrawled in Tagalog. A payslip photographed at an odd angle under the glare of a sari-sari store's fluorescent tube. An e-wallet transaction history that, upon closer inspection, doesn't quite square with what the borrower told the loan officer.
These are the raw materials of credit underwriting in Manila, Jakarta, and Mexico City—markets where open banking APIs remain more aspiration than reality. Manual review of such paperwork can drag on for days. Traditional optical character recognition chokes on real-world messiness. And somewhere in that gap between a borrower's digital exhaust and their actual creditworthiness, deals stall, fraud slips through, or perfectly good applicants get turned away.
Limcaoco and her co-founder, Rhea Malhotra, both Stanford-trained engineers, think they've found a way through.
What They Built
Their company, Kita—part of Y Combinator's Winter 2026 cohort—has developed a document intelligence platform designed to turn borrower paperwork into fraud-checked risk signals at speed. The system ingests more than 50 document types: bank statements, e-wallet records, payslips, utility bills, business income files. Out comes structured data—JSON, CSV, Excel—ready to feed directly into underwriting models. Kita's claim? It makes document-driven credit decisions 70 times faster than manual processes.
That's an audacious number. Whether it holds up at scale remains to be seen, but the early pitch has resonated. Limcaoco recently posted on LinkedIn that the company went "from zero to 93 KCR active in the Philippines and Indonesia" in three weeks, working with what she described as "some of the biggest banks and lenders." (KCR isn't publicly defined, though it appears to be a client or transaction metric the founders are tracking internally.)
The technical architecture leans on what Malhotra—who won Stanford's prestigious Firestone Medal for computer science in 2025—calls a "layered system led by vision-language models and computer vision." It's built to extract signals that legacy OCR misses entirely: handwritten ledgers, low-resolution smartphone photos, documents toggling between languages mid-page. The kinds of inputs that are routine in emerging markets but border on exotic for software trained on pristine U.S. mortgage files.
Why These Markets?
The Philippines rolled out its Open Finance Framework in 2021. Indonesia has made strides toward data portability. Mexico talks a good game about fintech infrastructure. But on the ground, the reality is patchier. Many lenders still rely on borrower-submitted documents as their primary source of truth—a vulnerability if those documents are tampered with, and an opportunity if you can extract better signals from them.
Limcaoco, who spent time building audio products at Apple after graduating from Stanford, has been logging miles: Manila, Jakarta, Bangkok, Singapore, and recently Mexico City. The thesis she and Malhotra are testing is straightforward. In markets where e-wallet penetration is high but standardized banking APIs are scarce, documents become the data layer. Kita positions itself as the connective tissue between messy paperwork and automated decision-making.
There's a pragmatism to the strategy. Open banking will arrive eventually—perhaps sooner than skeptics think—but until it does, someone has to parse the payslips.
The Edge They're Claiming
Kita's pitch rests on two differentiators, both of which sound plausible until proven at scale.
First: hyperlocalization. The system is tuned to recognize transaction patterns, document formats, and fraud signals specific to Southeast Asian and Latin American markets. That means understanding which e-wallets are commonly used in Metro Manila, what a legitimate payslip from a Jakarta logistics company looks like, or how Mexican utility bills format account numbers.
Second: a feedback loop that ties document-level signals to actual repayment outcomes. As lenders share portfolio performance data, the platform learns which extracted features correlate with default risk. It's the difference between simply digitizing a document and learning whether the patterns inside it actually matter.
The company also bakes in fraud and tamper detection—flagging altered PDFs, inconsistent metadata, or suspicious transaction patterns. That includes watching for gambling activity, a red flag in many emerging-market lending models. A live portal at portal.usekita.com lets users upload documents, run batch tests (up to 10,000 files on the enterprise tier), and pull results via API. There's also a Python SDK available on PyPI for developers who want programmatic access.
Crowded Space, Different Angle
Document intelligence for lending isn't a new category. Ocrolus has become something of a standard in U.S. mortgage and small business lending. Perfios operates deep in India and maintains a Philippines-specific product page. Inscribe has carved out a niche in fraud detection.
But Kita's founders argue those players optimize for mature markets—places where data is cleaner, API coverage is better, and the edge cases are less frequent. Their bet is that emerging-market lenders need a system purpose-built for messier inputs and market-specific risk signatures. Whether that thesis holds, or whether existing players simply expand downmarket, will become clearer as the competitive dynamics play out.
Malhotra's computer vision and robotics background anchors the technical architecture. Limcaoco's product experience and regional ties drive the go-to-market approach. The team is actively courting introductions to SME lenders, consumer finance companies, microfinance institutions, and traditional banks.
What Happens Next
Kita is marching toward Y Combinator's Demo Day on March 24. The company's launch video went live roughly a month ago; the portal is now publicly accessible. Pricing details beyond the free tier and an enterprise plan remain undisclosed—a common playbook for startups still experimenting with customer segmentation.
For now, the focus appears to be proving the 70x speed claim at scale and expanding document type coverage as new lenders come online. The real test will come when portfolio performance data starts flowing back. Can the system actually predict who pays and who doesn't? That's the question every lender will ask before moving from pilot to production.
In the meantime, Limcaoco's phone folder keeps growing. Another blurry payslip. Another handwritten ledger. Another e-wallet record that doesn't quite add up. Somewhere in that mess, she and Malhotra believe, is a business worth building.
