Applied scientist working at the intersection of math and machine learning.
I build things with machine learning and spend too much time thinking about the math behind them. Lately I've been digging into quantization theory, numerical linear algebra, and prediction markets.
If you want to work together or hire me for a project, email me at prateekchandrajha@gmail.com.
A complete training program for quant probability interviews. Nine anchor puzzles — from ants on a stick to the 100 prisoners — each teaching a named technique, followed by deep dives with 20+ harder variants at interview difficulty. Four standalone sections build martingales, order statistics, generating functions, and conditional expectation from the ground up. Covers 13 techniques, 30+ worked problems, and ends with a recognition flowchart and fact sheet for pattern-matching new problems on sight.
When does learning from data actually work? Every tool built from definitions — Markov, Hoeffding, Sauer–Shelah, symmetrization, KL divergence, Pinsker, Le Cam — proving the exact equivalence: a hypothesis class is learnable if and only if it has finite VC dimension. Upper bound, lower bound, and the complete Fundamental Theorem.
A complete, self-contained walkthrough of the easiest proof of the JL lemma. Every tool — expectation, MGFs, Markov, Chernoff, chi-squared concentration, union bound — built from definitions with examples, showing how small pieces snap together into one of the most useful results in high-dimensional geometry.
How random rotations, optimal quantizers, and a one-bit trick compress LLM memory by 6x while staying within 2.7x of the information-theoretic limit. Includes a puzzle about high-dimensional geometry.