How I Built an AI Agent That Checks if a Loan App Is RBI-Registered

TL;DR

Instant-loan apps promise money in minutes. But before you hand one your phone number, ID and bank details, there's a question most borrowers can't easily answer: is this app actually backed by a lender the RBI regulates?

So I built LoanGuard: Check Before You Borrow, a web-based loan-app risk intelligence platform for Indian borrowers.

The idea is simple: before borrowing through an instant-loan app, users should be able to investigate its publicly available information, understand potential risk indicators and see the evidence behind an assessment.

I'm especially proud that I integrated six TinyFish products into one workflow: TinySearch, TinyAgent, TinyFetch, TinyBrowser, TinyResearch and TinyMonitor.

In one live investigation of KreditBee, it found 27 public sources, read 10 pages and confirmed the app in the RBI directory, with Krazybee Services Limited as its regulated lending partner.

“A system should explain what it found, where it found it and what it could not verify.”

The problem

Instant-loan applications can be difficult for ordinary borrowers to evaluate.

The answers are public. They're just scattered across company websites, regulatory directories, public complaint pages and other sources, and nobody checks all of them when they urgently need a loan.

A borrower may not know:

• Whether an app appears in the RBI's public Digital Lending Apps directory.
• Which regulated lender or lending partner is associated with it.
• Whether grievance redressal details are disclosed.
• Whether interest-rate and business information is transparent.
• What public risk indicators have been reported.

Manually checking all these sources takes time and requires navigating different websites.

I wanted to bring these checks into one simple interface.

What I built

LoanGuard allows a user to enter a loan-app name and optionally provide its official website.

The system investigates available public information and generates a report containing:

• A preliminary risk score and verdict.
• Public-source findings and links.
• RBI directory verification status.
• Website transparency indicators.
• Browser-captured visual evidence.
• A cited research report.
• An optional RBI directory monitoring feature.

The frontend uses a real-time investigation timeline so users can see what each service is doing.

The result is designed to communicate not only potential risk, but also what could not be verified.

Where TinyFish fit

TinyFish powers the investigation pipeline through six products.

1. TinySearch

I use TinySearch to discover public information, websites, company details, complaints and other relevant sources.

2. TinyAgent

The agent navigates the RBI website and attempts to reach the public Digital Lending Apps directory. It checks for the app name and extracts the associated regulated lending entity when supported by directory evidence.

An important design decision was to distinguish a confirmed absence from an unsuccessful lookup. A timeout or access-denied response must never be interpreted as proof that an app is absent.

3. TinyFetch

TinyFetch retrieves publicly accessible web pages and extracts their content. LoanGuard uses this information to inspect grievance officer details, lending partner disclosures, business addresses and interest-rate information.

4. TinyBrowser

TinyBrowser creates a remote browser session, and I use Playwright over CDP to capture visual evidence from the target website.

5. TinyResearch

TinyResearch performs deeper multi-source analysis and generates a report with citations. I made it asynchronous so the initial investigation does not have to wait for a long-running research task.

6. TinyMonitor

TinyMonitor allows a user to create an on-demand watch on the RBI directory. It is optional and is not automatically created for every investigation.

What worked

Three implementation decisions were particularly valuable.

1. Parallel execution

Independent services are started concurrently, while TinyFetch and TinyBrowser start as soon as their required URLs are available. This avoids unnecessary sequential waiting.

2. Asynchronous result retrieval

TinyAgent and TinyResearch can continue after the initial report is displayed. Their actual results can subsequently update the investigation.

3. Evidence and status transparency

LoanGuard distinguishes confirmed findings from unavailable checks. It also keeps source evidence and browser screenshots connected to the report.

A further decision was to keep monitoring user-triggered rather than automatically creating recurring tasks.

A live check: KreditBee

I tested LoanGuard live on KreditBee, a popular instant personal loan app. TinyAgent found it in the RBI's Digital Lending Apps directory, with Krazybee Services Limited as its regulated lending partner. TinyFetch found grievance redressal details, a named lending partner and a registered business address on its website, and no complaint keywords turned up across the sources it inspected.

The verdict was lower risk, with high verification confidence.

Measurable outcome

In my recorded live KreditBee investigation:

These figures describe one recorded investigation, not a benchmark or guaranteed result for every app.

I have not yet conducted a statistically meaningful accuracy evaluation or a controlled time-saving benchmark. The score is a preliminary heuristic based on available evidence, not an official RBI rating.

What I'd do differently / Lessons learned

If I started again, I would design persistent asynchronous job tracking earlier.

TinyAgent and TinyResearch can take significantly longer than ordinary web requests. Keeping their run IDs, retrieving their results reliably and synchronizing late updates with the frontend required additional debugging.

I would also introduce a more extensive test suite for conflicting findings, source provenance and partial service failures from the beginning.

Recommendation for other builders

My biggest advice is to use each agent capability for a specific, meaningful job rather than integrating APIs just to increase the number of integrations.

Think carefully about what happens when a website is inaccessible, a request times out or two sources disagree.

In applications involving finance or other high-impact decisions, transparency is just as important as automation.

A system should explain what it found, where it found it and what it could not verify.

Try it / Links

The code is on GitHub, and the live app is free to try.

Live app → loanguard-five.vercel.app

GitHub repo → github.com/hemant2186/LoanGuard

Build Library page → tinyfish.ai/library/loanguard

Built at TinyHacks: Bounty Edition, GCET. More student builds are in the TinyFish Build Library, and if you're new to agents, start with build your first web agent in 15 minutes.

Want to build your own agent that checks the live web for you?

Sign up free → agent.tinyfish.ai

Docs → docs.tinyfish.ai

Open source Cookbook → github.com/tinyfish-io/tinyfish-cookbook

Building as a student? Join the TinyFish Student Program with your school email for $35 in free credits.