I build things where the math is the point, not the plumbing — and then I make sure you can see it happen, not just read about it.

BSc(Hons.) Data Science & AI, IIT Guwahati · PG Diploma in Applied Statistics, ISI Kolkata. Most of what's below started as "is this actually true?" and turned into a repo. One of them turned into an entry in the OEIS.

In August 2026, six terms I computed were accepted into OEIS A181671 — the count of Ramanujan primes below each power of ten. The sequence's EXTENSIONS section now reads:

a(18)-a(23) from Gaurav Vyas, Aug 15 2026

That pushes the table from 10¹⁷ out to 10²³, computed on a laptop and a desktop — no institute, no cluster — with a 128-bit segmented sieve, bracketed analytic tail bounds, and a machine-checkable certificate behind every term. The b-file now runs n = 1..23.

_{The definition, animated: π(x) − π(x/2) climbing, and each Ramanujan prime Rₙ lighting up at the last place the curve crosses level n — which is exactly what makes it Rₙ. Rendered fresh by CI.}

while (curious) {
    pick_an_equation();
    make_it_move();
    ask("why is this surprising?");
}

Mathematical Surprises → 250+ interactive visualizations — fractals, chaos, number theory, calculus — each one a real simulation you can drag sliders on, not a screenshot. Opens the same on your phone as on your laptop.

_{Nine rotating circles on the odd harmonics, tracing a square wave — one of the 250+ live pieces in Mathematical Surprises.}

Mathematics & number theory

Machine learning

Applied

The chess engine's self-play half did not work, and that's the interesting part. A 40-game match against its own frozen pretrained baseline scored 46.2% (95% CI 37.4–55.1) — no detectable improvement. The run before it was actively worse, and the reason took a while to find: the eval metric had been scoring one colour while the net alternated, so 159 logged evaluation points sat pinned at exactly 50% and carried no information at all. Mirror augmentation was also flipping board files without swapping the castling planes, quietly corrupting half of every minibatch.

The binding constraint turned out to be sample efficiency — 128 simulations is about 4.3 visits per legal move, against AlphaZero's ~27. All of it is written up in the repo's RESULTS.md, because a negative result you can reproduce is worth more than a positive one you can't.

I also build complete games solo — engine, art and audio. Both are zero-dependency and fully offline, with every sprite and sound generated at runtime.