When Not to Hint
Language-model query optimizers trained on one database underperform on a new one. I designed a low-cost pipeline to adapt them.
Language-model query optimizers trained on one database underperform on a new one. I designed a low-cost pipeline to adapt them.
A fine-tuning pipeline that teaches language models to write database query hints, cutting latency by 26% on unseen queries.
An all-in-one hiking backpack with solar charging, a GPS locator with emergency SOS, and a built-in water filter.
A desktop app that pairs PoseNet with formulas I designed to catch slouching in real time, with 97.3% accuracy.
A convolutional neural network that classifies heartbeat sounds as normal or abnormal with 98% accuracy, plus an Android app so people can screen their own heart sounds.
A new notation for the mridangam and a 21-layer CNN that identifies strokes with 94% accuracy, so recordings can be transcribed automatically.
A phone app that classifies a photo of trash into 7 categories with 97% accuracy and tells you which bin it goes in.
A lexer, parser, and tree-walking evaluator for a small language with loops, functions, and arrays.
Matrices, backpropagation, and a training loop with no libraries, trained to solve XOR.
A hackathon app that tracks how much water you drink and charts your habits.
A Discord bot for challenging friends to chess, with board images, takebacks, and draw offers.
A little-endian CPU emulator in C++ with almost 100 instructions, subroutines, and expansion ports.
A Python script that finds every valid set on the table.