hi, rahat here.

Most of my work sits at the same junction: using machine learning and NLP to detect things that are designed to look trustworthy but aren't: fraudulent job postings, malicious software packages, compromised IDE extensions, AI-generated text passing as human, and APIs that silently change behavior without breaking anything loudly enough to notice. The common thread isn't the application domain, it's the failure mode: systems that look correct under a naive evaluation and only reveal their flaws under adversarial scrutiny.

That's also how I evaluate my own work. Every one of these projects has gone through multiple rounds of self-directed adversarial review before ever reaching a peer reviewer, hunting specifically for train/test leakage, confounded baselines, and claims that don't survive a harder look. This has repeatedly caught real problems in my own results: near-duplicate leakage inflating a malware classifier's reported accuracy, a headline finding that reversed once evaluated with proper statistical correction, an implementation that quietly never ran the correction it claimed to apply. I'd rather find these before submission than after.

What I want to spend a PhD on is making that kind of self-correcting rigor less dependent on any one researcher noticing by hand: better evaluation protocols, more honest baselines, and detection systems for software engineering and ML that are held to the same adversarial standard they're meant to enforce on their targets.