Keshav Sundararaman

End-to-end ML pipeline predicting abnormal next-day price moves across 505 S&P 500 stocks (619K+ daily OHLCV rows spanning 5 years), engineering 12 no-lookahead features 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. Random Forest performed best (test AUC 0.72, F1 0.31 at a 2% move threshold); used feature importance and L1 sparsification to identify volatility and price-range features as the strongest signals.

Built a 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. Designed the architecture first, then built it with Claude as a dev partner. Shipped v1.1 within 24 hours of launch adding support for size-and-color variant grids and bundle listings.

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.

Amrita Vishwa Vidyapeetham × Sony India

Built automated network scripts to continuously ingest raw IoT sensor readings into structured tabular datasets. Cleaned and validated multi-sensor experimental logs in Pandas to eliminate sensor noise, drift, and missing entries. Analyzed food drying behavior across sensor runs and presented analytical findings and Matplotlib charts directly to Sony representatives and project leads.