A loan officer in Manila confronts a familiar sight: bank statements from three institutions, none formatted alike. Payslips on company letterhead that could be real or fabricated. Utility bills. Tax documents from the Bureau of Internal Revenue. Perhaps a credit report, if the borrower has one. The verification process—checking transaction histories, cross-referencing employment claims, watching for signs of tampering—can stretch across days.
Kita, a two-person team fresh out of Y Combinator, believes that timeline should collapse to minutes.
The San Francisco-based startup is building software that uses vision-language models to transform those messy piles of financial documents into structured credit signals. The initial focus is markets where banking infrastructure remains fragmented or incomplete: the Philippines, Indonesia, Mexico. Places where "credit histories live in documents," as the founders put it, rather than in tidy APIs.
It's a specific problem, though not necessarily a small one. And the solution, if it works, carries implications for how millions of people in emerging economies access capital.
Messy Documents, Structured Signals
The core product, called Kita Capture, ingests the financial detritus that defines creditworthiness in these markets. Bank statements, yes, but also e-wallet transaction histories, audited financials, credit bureau reports from local agencies like the Philippines' Credit Information Corporation, government-issued tax forms (BIR 2303s, 2307s, income tax returns), and national IDs. The system validates data across documents, flags evidence of tampering, categorizes transactions—including, notably, gambling spend—and extracts underwriting signals.
According to the company's Y Combinator Launch page, posted in late January, Kita makes "document-driven decisions 70x faster, with built-in fraud checks." The product slots into existing underwriting workflows, either via web portal or API, and includes what the company calls a learning loop: the system refines its signal extraction based on actual loan performance over time.
In mid-February, co-founder Rhea Malhotra released an official Python SDK. The documentation lists support for over 50 document types—everything from bank statements and paystubs to Secretary's Certificates and proof of billing. Outputs include transaction-level data, employment verification, fraud scores, and extracted underwriting signals, serialized to JSON, CSV, or Excel.
The technical approach centers on computer vision rather than simple OCR. The system doesn't just read documents; it learns relationships between document-level signals and repayment outcomes. In a February LinkedIn post, the company detailed capabilities that include cross-document validation and what it calls "analyst decision support." How much of the origination process this actually automates remains, for now, a bit opaque. Specific benchmarks beyond that "70x faster" claim have not been made public.
The Infrastructure Gap
Kita's initial geography is deliberate. The Philippines, Indonesia, and Mexico represent markets where open banking APIs and comprehensive credit bureaus exist in some form but remain nascent or fragmented. The Philippines offers a useful case study.
The Bangko Sentral ng Pilipinas and the International Finance Corporation launched an Open Finance PH pilot in 2023, followed by a hackathon the next year and ongoing trials through 2025. Progress is real, but phased. Standardized, interoperable data access—the kind that companies like Plaid have turned into infrastructure in the U.S.—remains a work in progress.
Credit data availability is improving, though coverage gaps persist. The Credit Information Corporation generated 10 million credit reports in 2024, according to Philstar. That figure jumped to 27 million in 2025, per BusinessWorld. Rapid growth, certainly, yet millions of potential borrowers still lack formal credit histories. In that vacuum, documents become the default vehicle for risk assessment.
For lenders, this creates friction. For Kita, it creates an opening. Where permissioned bank data flows through aggregators in New York or London, lenders in Manila or Jakarta are often working from PDFs sent via WhatsApp. That's the workflow Kita is trying to replace.
The architecture appears designed for localization rather than generalization. The company emphasizes "hyperlocalized signals" for each market, suggesting the models are trained or fine-tuned on region-specific document formats, transaction patterns, and fraud vectors. A customization engine allows individual lenders to adjust signal extraction based on their own portfolio data. Whether that flexibility translates into measurable improvements in approval speed or default rates is the bet.
Early Traction, Unverified Claims
Kita was founded in 2025 and entered Y Combinator's Winter 2026 batch, securing the accelerator's standard deal: $500,000 total investment, comprising $125,000 for 7% equity on a post-money SAFE, plus $375,000 on an uncapped most-favored-nation SAFE. The company is working with Ankit Gupta as its primary YC partner.
In February, co-founder Carmel Limcaoco posted on LinkedIn that the team would be on the ground in Manila, Singapore, Jakarta, and Bangkok over three weeks to deploy with SME lenders, consumer finance companies, microfinance institutions, and banks. A later update mentioned the company was selected to present at YC Launch Live and claimed traction: "In three weeks, went from 0 to 93 KCR active in the Philippines and Indonesia, working with some of the biggest banks and lenders."
The company has not publicly defined "KCR," and no lender partnerships have been disclosed by name. Take the claim with appropriate skepticism, perhaps, but the founders are logging miles.
Kita launched on Product Hunt around mid-March. The listing was succinct: "Turn documents into signals for lenders."
The Founders
Carmel Limcaoco, CEO and co-founder, is from Manila and studied Symbolic Systems and Music at Stanford, where she also pursued a master's in computer science. She spent three summers working on audio and music products at Apple and previously co-founded DAHA, a Stanford marketplace. Limcaoco has been vocal on LinkedIn about the company's field work, posting updates from Southeast Asia and Mexico City with the kind of energy that suggests she believes deeply in the problem.
Rhea Malhotra, CTO and co-founder, holds a BS and MS in computer science from Stanford, where she was awarded the Firestone Medal for research. She deferred an incoming PhD in Computer Vision and Robotics at Princeton to build Kita. Malhotra maintains the company's Python SDK and handles the technical infrastructure.
The Y Combinator directory lists the team size as two. No other employees are mentioned.
A Fragmented Competitive Landscape
Document automation for lending isn't new. Ocrolus offers AI-powered processing with fraud detection for bank statements, pay stubs, and tax forms—largely for North American lenders. Google's Document AI includes lending-specific processors for bank statements and payslips. AWS offers Textract Lending Analysis workflows. Niche players like LendAPI and TryDocu serve similar use cases, primarily in the U.S. and India.
What appears less crowded is purpose-built infrastructure for the document formats, languages, and fraud patterns specific to Southeast Asia and Latin America. In the Philippines, FinScore focuses on telco data for credit scoring—a different alternative data vector—and recently announced a partnership with DITO to reach 100% mobile subscriber coverage. TransUnion and FICO have piloted alternative scoring models in Kenya and other emerging markets, underscoring global demand for non-traditional credit signals.
Kita's bet is that vision models, trained on localized documents and continuously refined by repayment outcomes, can bridge the gap more effectively than generic OCR or U.S.-focused products retrofitted for new geographies. It's a reasonable thesis. Whether it holds depends on execution: model accuracy, fraud detection rates, integration friction, and whether lenders see measurable improvements in the metrics that matter.
For now, the product is live, the SDK is shipping, and the founders are crisscrossing Southeast Asia. The rest is a question of scale.
