Raunak Sood

A test for data leakage that needs no answer key. You say which tables each feature may read; it changes something the code must be blind to, like one table's timestamps, and re-runs. Anything that moves was reading something it shouldn't.

Result. A real leak found this way had inflated a stock-picking result by 59%. Released on PyPI; on five public datasets it catches 8 of 9 planted leaks with no false alarms, and the ninth is invisible to it by construction.

Score each step of an AI's reasoning by freezing the search there and re-running it ten times. The share of re-runs that still succeed is what the step was worth.

Result. A per-step map over 22,350 re-runs. The AI's top-ranked move led to a solution 16% of the time; lower-ranked ones about 40%. A five-author course project, built and run collaboratively.

Measure how extraction recall changes with the number of items to find by building the documents from verbatim sentences, so the answer key is exact and free.

Result. In a one-model pilot, free-form recall falls 0.214 from 1 to 16 items and schema-constrained recall does not. The gap rests on the shortest lists, so the schema is not the cost; whether it is the cure is open.

A dated map of which companies name which others in their 10-Ks, every edge quoting its sentence, built with a model on a laptop.

Result. 2,572 relationships over 220 firms, released with its coverage ceiling attached: 47 names per day against the 97 the return test needs. The reason is disclosure law, which makes customers speak and suppliers silent.

An app for photographers, models and stylists to find each other: swipe, match only when both say yes, then talk.

Result. Live on the App Store and the web, two clients against one Postgres, with the feed ranked by a query and mutual consent enforced by the database. No real user base yet.