Command Center

How NIFFLER has done on a fixed set of 500 made-up orders. See Getting Started for where this data comes from.

Revenue at Risk

₹4,77,400

200 failed orders found

34 were left alone on purpose, as a comparison group to show what NIFFLER actually added

Revenue in Recovery

₹2,72,673

₹15,195: confirmed
₹2,57,478: waiting on payment links

Confirmed means the money is in. The rest are real payment links that have been sent, but these customers are made up, so nobody is ever going to click them.

Recovery Rate

3.6%

-2.6 pts vs control

Compared with 6.2% in the group left alone

This only counts money that came in. Since made-up customers never pay a link, the ₹2,57,478 above can never be counted here. That is a limit of the test data, not of the agent.

Recovery Cases

166

34 more left alone for comparison

All done

Recovery Actions

127

5 payments taken, 122 links sent

Escalations

Make your own failed paymentOne real test order. Fail it on purpose and watch the agent take it apart. Opens in a new tab.Watch a case run livePick one from the dataset and follow every step as it happens.

8 cases created by visitors3 recovered (₹1,098), 5 waiting on a payment linkSee them

These came from real Razorpay test payments people made on this site. They are kept out of the numbers above on purpose, so the measured batch stays fixed and comparable.

Would this survive lakhs of orders?

Honest answer: it has been run on 200 orders, not lakhs, so anything else would be a guess. But the shape of the problem is clear enough to say where it would bend.

Most of it never touches AI

Finding the failed orders is plain code over the database. Checking the rules is plain code too. Both are fast at any size. Only the diagnosis needs a model, and only for orders that are worth chasing in the first place.

The model is the bottleneck

Each case costs several AI calls, and cases are handled one after another today. That was a choice for correctness, not speed. The real ceiling I hit was the free tier's daily token limit, not the code.

Three quarters could skip the model

The rules comparison shows plain rules reach the same answer 73% of the time. Send the easy majority through rules and spend the model only on the rest, and the AI cost drops by roughly three quarters.

What is genuinely missing for that scale: cases are processed one at a time rather than in parallel, the run happens inside the web request instead of a proper queue, and none of it has been load tested. Those are real gaps, not ones I am going to pretend are solved.