Punarnava

I asked the same AI model the same question twice. It gave two different answers, and the more confident one was probably wrong. I'm building AR3 for the Nebius × NVIDIA Global AI Hackathon: a layer that works out why a UPI or card payment failed and recommends a safe next step. Last night I ran a smoke test on NVIDIA Nemotron 3 Nano through Nebius Token Factory. One failed UPI payment: error U30, "debit failed." Temperature 0. Only one setting changed: reasoning on or off. Reasoning on: "temporary gateway glitch," confidence 0.85, 14 s, 405 tokens. Reasoning off: "hard decline, never retry," confidence 0.95, 2.5 s, 145 tokens. 5.6x faster, 2.8x cheaper, and more sure of itself. But "debit failed" is one of the vaguest errors in UPI. Calling it a permanent decline means giving up on a payment that might have gone through on a retry. What I took from it: 1. A model's confidence score is not evidence. 2. Reasoning costs time and tokens, so it's worth measuring where it actually changes the answer. 3. The model shouldn't make the final call on money. That's how AR3 is built. Known error codes go to a rule table. Only unclear ones go to Nemotron, fast model first and Ultra when it's unsure. Then independent safety checks decide. AI diagnoses, rules decide, and AR3 never moves money. One example isn't a conclusion. Next: running this across the full test set to measure accuracy, speed and cost for each setting. Repo (work in progress, hackathon deadline: Oct 30): github.com/punar4/ar3 @nebiusai @nebiustf @Nebius_Academy @NVIDIAAI #Nemotron #BuildInPublic

I'm on a saving spree these days, everytime I open X, someone's giving a roadmap on how to become ai engineering or ML engineer or backend. But saving itself won't be enough, so I'm planning to build a 'resource archive' so that I can store it there for now and find it when I am actually looking for it. Might as well create a 'resource search', to search on all platforms for resources at once.

lets see if this apporach works: @rananava Hi Raoul, I'm Punarnava. I'm building AR3, an open-source layer that works out why a UPI or card payment failed, especially vague errors like UPI's "debit failed", and recommends a safe next step. AI diagnoses, fixed rules have the final say, and it never moves money. Before I decide who it's for, I want to hear from people who deal with this for real. One question: What does Navana.ai do today when a UPI or card payment fails? A two-line reply would help a lot. I won't quote you anywhere without asking first. Thanks, Punarnava

Rest at the end, not in the middle.

Don't give advice, become an example. ~R N Kao

AR3 Chapter 0 - The rebuild begins. My project scored 4.5/10 in a code review. The review was fair, so I’m rebuilding it. AR3 is a decision engine for failed payments. When a UPI or card payment fails, it’s built to work out why and pick a safe next step: retry through another acquirer, retry later, prompt the customer, or stop. The design: rules handle known error codes, an LLM diagnoses the messy ones, and independent safety checks have the final say. v1 didn’t live up to that. The README claimed more than the code did, there were no tests, and when the LLM call failed, the code quietly defaulted to rerouting the payment instead of stopping. That’s the exact failure the design was meant to prevent. v2, in order: honest docs, real decision logic, validated LLM output, tests, then a live demo by Oct 31. Biggest lesson so far: a demo that works once is not a system you can trust. I’ll share the demo here when it’s live. If you’re hiring SDE or AI interns, I’d love your feedback on it.

Simply Put #1 Key terms and concepts of my post, simplified -

Simply Put #0

I did nothing today.