Keshav Sundararaman

Random Forest reached test AUC 0.72 (F1 0.31 at a 2% move threshold) predicting abnormal next-day price moves across 505 S&P 500 stocks. End-to-end ML pipeline over 619K+ daily OHLCV rows spanning 5 years, with 12 no-lookahead features engineered from price and volume history. Used strict time-based train/validation/test splits to prevent data leakage and tuned hyperparameters on validation F1 across four model families. Feature importance and L1 sparsification identified volatility and price-range features as the strongest signals.

Live Chrome extension that scans all variants on an Amazon product page and finds the lowest price across every condition: New, Like New, Very Good, Good, and Acceptable. Shipped v1.1 within 24 hours of launch, adding support for size-and-color variant grids and bundle listings. Designed the architecture first, then built it with Claude as a dev partner.

Picks 2.19 of 3 podium drivers right per race, against 2.07 for “top 3 on the grid”. Predicts the podium of each 2026 Formula 1 race from qualifying results, recent form, practice pace and weather. For every race it tests 7 scikit-learn models on 43 past races, training on earlier races only, and keeps the best one. Each race has its own notebook and a public scorecard, misses included.

A private tracker for LinkedIn connection requests, built to be updated by Claude in one line. It has a blank tracker page, a Claude skill that adds and updates rows with minimal tokens, and a guide to writing short, specific connect notes. Nothing is automated; you send requests yourself and it tracks status and follow-up timing.

Comparing three pretrained vision backbones (ResNet-50, CLIP, DINOv2) as frozen feature extractors on EuroSAT, measuring how classification accuracy scales with labels-per-class to find where a trained probe overtakes zero-shot transfer. Pipeline validated end-to-end at reduced scale.