aaron chakraborty · ai/ml systems architectoffensive security researcher · b.tech cse @ kiit · graduating 2027
← that second line is clickable. everything changes. ← so is the first one. two sides, same person.
A 567M-parameter MoE transformer, trained from scratch on an A100. My architecture, my calls, my bugs. I can explain every layer because I had to debug every layer. Prompt injection, jailbreaks, data extraction. I found an XSS in a production AI chatbot, reported it quietly, and the company fixed it. Everyone won.
//
That claim up there? Here's the log. And here's what breaking looks like.
Every line below is from a real training run. Real loss curve, real step count, real hardware. A real scan, ending in a real bug on a real production system. Reported and fixed.
click the pane to skip the typing · toggle BUILD/BREAK in the nav to see the other half
If you don't read code, here's the translation: that box shows a language model, the same family of thing behind ChatGPT, being trained from zero. Not customised, not plugged into an API. Started from an empty file, and it learned. In plain words: companies ship AI chatbots to millions of users. I found a security hole in one, told the company privately, and they fixed it. Nobody got hurt. That's the whole point.
//
Where do you
want to start? Pick your
entry point.
// 01
The honest versionThe guy on the other side
SSB rank 83, a lab full of people I call family, and why I picked the hard road.Know your adversary. Background, motivations, and the day the army almost got me.
$ cat about.md →
// 02
Claims & receiptsThe operations log
DBS, SwiftSafe, 20+ VAPTs, and one real bug in production. Every claim carries proof.20+ assessments, one production XSS a real company had to patch, all of it legal.
$ cat work.log →
// 03
The systemsThe arsenal
VantaLLM, Argus, ReconForge. Built end-to-end, tested until boring.Argus gates AI agents before they ship. ReconForge maps attack surface. VantaLLM is the practice target.
$ ls projects/ →
// 04
From the trenchesCrash reports
I post the crashes along with the wins. The OOM errors nobody warns you about.Things going wrong, documented while still annoyed. Failure is the best teacher I ever had.
$ tail -f trenches.md →
// 05
In their wordsDon't take my word for it
What people say after working with me, live, named, unedited. Add your line.Independent sources, real names, zero curation. Verify me against people who were there.
$ cat wall.log →
// 06
Run the scanOpen a channel
Argus evaluates AI agents before deployment. This copy runs on me. Try it.Every way to reach me, plus a scanner that reports on its own author. Run it.
$ scan me →
