Abhijit S

Aspiring developer sharing my journey and projects as i build them

[Abhijit S @Abhijit_svs_dev]

Stilll after almost 15 days of constant effort couldn't get selected to the Razor pay internship
It would make my first internship rejection but learnt something of worth while ig
Hope things turn around next time
#razorpay

[Abhijit S @Abhijit_svs_dev]

Completed the product complety and fully
Then took another 5 days to compete the video presentation for the product

[Abhijit S @Abhijit_svs_dev]

About building an AI Payment Recovery Agent for Razorpay.

The initial prototype was simple. Then real edge cases hit: model collapse, discount abuse, and audit gaps.

300 scenarios and 3 recovery loops later, here is what actually happened 🧵👇

[Abhijit S @Abhijit_svs_dev]

Building a payment recovery agent for Razorpay.

The LLM recommends. Policy checks. Executor acts.

Things broke today: bad data, model limitations, and an audit gap.

Fixed them. 100 payments now run with 0 policy violations.

#BuildInPublic #AI @Razorpay

[Abhijit S @Abhijit_svs_dev]

@Suu766 It's only for US students

[Abhijit S @Abhijit_svs_dev]

If an AI is moving revenue, you can’t just hope it logged everything honestly.

I added two simple SQLite triggers directly on the audit table: • Block all updates • Block all deletes

Even if my Python backend crashes or bugs out, the paper trail is physically unchangeable.

[Abhijit S @Abhijit_svs_dev]

I tried running a 1B model (gemma3:1b) locally to save latency.

It collapsed. On conditional logic 1B returned the SAME action for every payment category regardless of context.
So I Upgraded to mistral latest 7B: 100/100 inferences
0 fallbacks Multi-branch reasoning restored

[Abhijit S @Abhijit_svs_dev]

2/14 Standard webhooks log payment failures, but nobody decides what happens next.
Traditional software logs. AI decides.
Our architecture rule:
🤖 LLM Recommends
🛡️ Policy Engine Checks
⚡ Action Executor Acts
If AI and Policy disagree, Policy wins EVERY single time.

[Abhijit S @Abhijit_svs_dev]

Building a payment recovery agent for Razorpay.

The LLM recommends. Policy checks. Executor acts.

Things broke today: bad data, model limitations, and an audit gap.

Fixed them. 100 payments now run with 0 policy violations.

#BuildInPublic #AI @Razorpay

[Abhijit S @Abhijit_svs_dev]

Finally active on X :)

Into AI, GenAI, Python, DSA, AWS and building random projects.

Looking for people who are into the same stuff.

If you are, let's connect.

#BuildInPublic #AI #Coding

[Abhijit S @Abhijit_svs_dev]

@luannguyenbkit Let's connect

[Abhijit S @Abhijit_svs_dev]

@Argona0x The matrix is here

[Abhijit S @Abhijit_svs_dev]

@devops_prashant YouTube premium is always the worth for the money