Hey, I am
I write economics research that goes out to peer review, and I ship production software that takes real money from strangers. Usually in the same week. One paper is with reviewers right now. One product is taking payments while you read this.
Memory, spuriousness, and a model-free entropy diagnostic in homogeneous-good Bertrand competition.
Two pricing algorithms that cannot communicate still settle on high prices and punish each other for undercutting. The paper shows the margin is memory, not sophistication. The simpler stateless learner colludes far more often but weakly and spuriously, while the stateful learner colludes rarely and genuinely, and it introduces a diagnostic that tells those two apart.
That distinction matters because the law hunts for an agreement, and here there was never one to find.
Written, coded and submitted alone, with no institution behind it. It cleared editorial screening and is with reviewers now.
A pip-installable, Gymnasium-style suite of economics environments for reinforcement learning, so an off-the-shelf agent can run against a market model without being rewritten. Packaged, tested, with a software paper alongside it.
Computational attacks on small Ramsey-type problems and no-three-in-line configurations. Local search, symmetry reduction, and exhaustive verification across very large candidate spaces. Built for correctness first, because a combinatorial claim is either right or it is worthless.
AI companion safety. Surprise detection, metacognitive monitoring, and the harder question underneath both: what a system should do when its read of someone's situation is correct and it still has to answer well.
A consumer product I designed, built, deployed and still run. Next.js 16 and React 19, Razorpay checkout with server-side HMAC-SHA256 signature verification so an order is never trusted from the client, and a Cloudinary pipeline for customer uploads. Strangers customise, pay, and receive their page with nobody in the loop.
Tailored sites for funded startups, each one researched, designed and deployed to a live URL. Several are bilingual English and Arabic with full right-to-left: mirrored layout, direction-aware components, locale-correct typography. Unsolicited work, and labelled as concepts on the pages themselves.
A commissioned teardown of a shipped LLM feature. Nineteen documented defects across prompt handling, output validation and failure behaviour, plus adversarial testing against eight prompt-injection and jailbreak vectors, each with reproduction steps. Written for the engineers who had to fix it, not as a list of findings to admire.
A crime that has always needed two people in a room, committed by two programs with no way to speak to each other. The whole argument of the paper, with no equations and no jargon.
Read →How to tell whether an AI companion is helping you or only making you feel better. One question, and you ask it after you close the app rather than during.
Read on obvix.io ↗I work remotely from India and invoice as a contractor, so there is no visa and no paperwork on your side. Contract or full time, whichever is easier to approve.