MollyAI

Inspiration

Be honest: when did you last check your credit card bill line by line? Most of us open the app, see the total, and pay it.

We did the same, until we looked at the numbers. 6 in 10 Indian online banking users have been hit with hidden charges that show up only later [1]. 1 in 5 people are actually afraid to open their statement [2]. And 42% forgot they were still paying for a subscription [3].

Checking a bill by hand is not easy either. Interest, late fee slabs, GST, cash advance fees, forex markups and EMIs each follow their own rule. And pasting your statement into a chatbot means handing over your name, card number and address to a model that might just guess the maths.

So we asked ourselves: what if something checked the bill for you, properly, without ever knowing who you are?

What it does

You upload your statement (password protected PDFs work too) or pick one of our 15 samples. In under a minute, MollyAI:

recalculates every fee, interest charge, EMI and total against 37 rules from the card's terms
compares it with earlier months to catch duplicate charges, price hikes and missing refunds
shows proof for every problem: what it should be, what was charged, and the exact lines behind it
explains it all in plain English and tells you what you should actually pay
writes a dispute draft for you in one click

If nothing is wrong, it simply tells you you're good to pay.

How we built it

We followed one rule throughout: code decides, the AI explains.

Every amount and alert comes from plain Python rules built from the card's terms. NVIDIA Nemotron 3 Super, running on Nebius Token Factory, only does the human part: reviewing alerts, summarising the month, answering questions and drafting disputes. It never does the maths.

Before anything reaches the AI, names, card numbers and other personal details are swapped for tokens like NAME_1. Guardrails block any request that still carries personal data and reject any reply with an amount we never sent. Uploaded files are deleted right after reading.

We also built our own evaluator to check every report for correct numbers, real evidence and privacy leaks, because "it looks right" was not good enough for us.

Challenges we ran into

The hardest part was not finding errors, it was knowing what is not an error. A late fee after a payment made within the grace period can look wrong when it is perfectly fine. We ended up writing these "allowed" cases as carefully as the error rules.

Our evaluator also needed fixing. At first it confused card number digits with amounts and flagged things that were correct. And running a model locally was slow, so we moved to Nemotron on Nebius, which made it fast enough for anyone to try live.

Accomplishments that we're proud of

12 out of 12 planted errors caught, with zero false alarms
100 out of 100 errors caught in our stress test
not a single piece of personal data reached the AI
zero made-up amounts in AI replies
about 0.001 USD to check one statement

Our favourite part is the Error Lab. Judges can plant their own mistakes in a clean statement and see if MollyAI catches them.

What we learned

AI is brilliant at explaining things, but your money should be calculated by code. Privacy has to be built into the design, not added as a line in a prompt. And testing your work deserves as much effort as building it.

What's next for MollyAI

Right now MollyAI reads our sample statement format, which closely follows a real Indian credit card statement. Next, we want to support real bank layouts, add rulebooks for more cards, and alert you the moment a new statement arrives.

All demo statements are dummy data, and the rulebook is our reading of a sample card's terms, not official bank policy. Sources 2 and 3 are US surveys.

Sources: