Demo
Armaan KapoorMarch 20, 20265 min read

The tagging problem

How institutional knowledge stops being oral tradition and starts being infrastructure.

Every transaction in a bank statement has a description, an amount, and a date. That's all the bank gives you. What the transaction actually means for underwriting is a different question entirely. Is this debit a loan payment or a vendor expense. Is this credit revenue or a disbursement cycling through. Is this fee an NSF or an overdraft. Is this transfer between the business's own accounts or money leaving the business.

The descriptions don't answer these questions. They're truncated ACH strings full of routing metadata and confirmation codes. The same funder shows up as three different descriptions across three different banks. The same description means different things depending on the amount. The same transfer is internal or external depending on whether the destination account belongs to the business.

Underwriters have always resolved this ambiguity through pattern recognition built over years of looking at statements. They recognize funder names. They recognize payment shapes. They know that daily debits at a consistent amount are almost certainly debt service. They know that a large credit followed by periodic pulls is a funded position. This knowledge is real and it is valuable and it has never been captured anywhere. It lives in the underwriter's head and it leaves when they leave.

We built a system that captures this knowledge and makes it persistent.


Transactions compress into semantic groups. Similar descriptions collapse together so that one classification decision covers an entire family of related transactions. Three classification engines run in parallel. One catches unambiguous patterns deterministically. Checks, wires, peer-to-peer transfers, fees. A second classifies groups into loan types. Merchant cash advance, bank loan, factoring, lease, auto, mortgage. A third identifies activity types. Internal transfers, owner transactions, payment processors, reversals.

Each org on the platform maintains what we call a funder registry. This is a living database of every lender and funder the org has encountered. Names, aliases, transaction description patterns, contact information. When the system recognizes a group of transactions as matching a known funder, it links them. When it encounters a funder it hasn't seen before, the underwriter identifies it and the registry learns. That alias is now permanent. Every future deal that contains that description pattern gets matched automatically.


The underwriter reviews the output and corrects where needed. Retag transactions in bulk. Drag a position from one loan type to another and every transaction underneath retags with it. Merge positions the system incorrectly split. Create new positions for activity the system missed. Exclude a document or an account from the analytics. Every edit propagates instantly. Revenue recalculates. The debt stack updates. The spreadsheet regenerates. Screening rules re-evaluate.

Every correction is a lesson. 59,000 corrections across the platform and counting. Each one made by an underwriter who saw something the system didn't. Each one persisted into the org's funder registry. Each one making the next parse more accurate than the last.

The system holds the volume and the history. The underwriter holds the judgment. And every time the underwriter exercises that judgment, the system absorbs it and carries it forward. This is how institutional knowledge stops being oral tradition and starts being infrastructure.

Keep reading

View all