Not long ago, getting a consumer loan could take hours or even days. Most steps ran sequentially and required human involvement: data verification, credit history analysis, risk assessment, fraud detection.
Today the picture is changing fast. AI and automated rules engines make it possible to move much of the underwriting process into a digital environment, where a decision on a loan application can be made in a matter of minutes.
Taras Nahai, Senior Developer at Devinware LLC, explains how this infrastructure works, what technical challenges lie behind automated lending, and why AI is becoming an integral part of financial products.
Taras designs and develops a rules engine for automated credit underwriting, integrated with several US credit bureaus (TransUnion, Clarity, MicroBilt), allowing application decisions to be made in minutes rather than through lengthy manual review.
He currently works as a Senior Developer at Devinware LLC, his own consulting company in North Carolina.
Nahai says that before his rules engine was introduced, manual review of a single loan application took up to an hour. Now it takes 10–15 minutes. The system also lets the company control the cost of processing each application. The price of every check is recorded in dollars, and it is possible to see how much the analysis costs across different data providers, which makes it possible to optimize spending at every stage. For example, if an applicant fails the basic checks, the system can reject the request after evaluations costing about $0.60. A more expensive scenario, costing $4.93, is triggered only for applicants who have passed the earlier stages and reached income verification. In this way, the system makes lending decisions automatically while managing costs: expensive checks are applied only where they are truly needed.
The AI does not calculate financial metrics on its own; it relies on deterministic tools and fixed formulas applied to verified data. Every conclusion retains its reasoning, the prompt version and the full set of tool parameters, so it can be reproduced and audited. In addition, a point-in-time data snapshot captures the state of all key signals at the moment the decision was made, so the historical record remains unchanged.
The reason for a rejection is built from the specific rules and indicators that influenced the decision, so the borrower can be told clearly which factors led to the outcome. The system records the specific rules and indicators behind each decision, so an applicant can be given the reasons for the outcome – the level of detail that ECOA / Regulation B and CFPB Circular 2022-03 address.
Taras works on optimizing high-load systems: processing message queues in the millions, cutting production query execution times from hours to minutes, and archiving hundreds of gigabytes of data with zero downtime. The system processes around 4.7 million messages per month, and that figure continues to grow. On average, that is roughly 152,000 messages per day. The software handles a large daily data flow and processes messages automatically, without the need to handle each one manually.
The expert explains how production queries were sped up: “This concerns the export of a large leads database. Previously, the standard approach required loading the entire file into memory before saving it. Because of this, large datasets, for example 3 million rows, exceeded the available memory. On top of that, processing the data in chunks through the old mechanism got slower and slower as the table grew. As a result, an export could stall after about 20 minutes without ever finishing the file. I changed the approach: now the data is written to CSV gradually, in small portions, without having to hold the whole file in memory. I also modified the way data is selected, so processing no longer slows down as the database grows. In addition, I optimized statistics calculation and memory handling. As a result, the export has become stable and predictable: roughly 20 minutes per 1 million rows, while memory consumption stays practically unchanged regardless of data volume. This makes it possible to export huge datasets, effectively any number of rows, without the process crashing.”
To build a fuller picture of a borrower’s financial situation, the system processes alternative data such as bank transactions, income information from paystubs, utility payments, and details about the phone, IP address and device. This relies on services including Plaid, Edge, Clarity, MicroBilt, TransUnion and IPQualityScore, while income documents are analyzed using AI.
Taras says that a typical borrower who previously would not have received a loan but now does is someone with a stable job and regular income, but whose credit history or score does not reflect their real ability to pay. “The system analyzes the client’s bank transactions, including the regularity of salary deposits and the amount remaining after mandatory payments. The available loan amount is determined in advance based on these indicators and the client’s actual ability to repay the loan,” he says.
AI is gradually transforming consumer lending. Underwriting is turning from a long, manual process into a fast, automated and more controllable system. Rules, verified financial data, alternative signals and AI shorten decision times and allow a more accurate assessment of a borrower’s true solvency. The transparency and controllability of such systems make it possible to reproduce, verify and explain every outcome.
