this site has two modes.

BUILD, the engineer who trains models.
BREAK, the one who attacks them.

flip the switch ↑ and everything re-frames.

// hire · the 60-second version

In a hurry?
Here's the whole case.

Aaron Chakraborty. AI/ML engineer and security researcher: I build systems from scratch, then test how they fail.

B.Tech CSE @ KIIT · CGPA 9.42 · graduating 2027 · currently Technology Apprentice @ DBS Tech India (SEED)
open to: AI/ML engineering · security engineering · placement season 2026-27

three things that are true and verifiable

  1. 1. Trained a 567,068,416-parameter MoE transformer from scratch on a single A100: my architecture, 30,000 steps, 86/86 tests. case study
  2. 2. Found and responsibly disclosed a reflected XSS in a production AI chatbot (Yatra.com); vendor acknowledged and fixed it. the disclosure
  3. 3. AIR 45, top 0.90%, at Pentathon 2025, a 48-hour national CTF, playing mostly solo. Plus eJPTv2 at 91%.

two systems worth your time

VantaLLM →

567M-param MoE LLM, from-scratch training framework. Scales 567M → 125B by config.

Project Argus →

Two-layer AI-agent security platform: release gate + runtime gateway. Public repo.
download resume ↓ email me → linkedin →

got more than 60 seconds? the full site is a much better story.