Michael Ku Jr

A RAG system over long documents, built evaluation-first: 94.3% answer accuracy against a 38.6% no-retrieval baseline on a four-document corpus, behind a CI gate that fails the build when quality drops.

Detecting manipulated photographs from camera and compression physics rather than a trained real/fake classifier. Measured on 112 hand-made forgeries and their matched originals.

A local-first agent that ingests AI/ML videos into a concept-and-tool knowledge base, answers with clickable timestamps, and tracks what I already know so it stops recommending it. Runs on CPU at $0.

A planner that teaches people building apps with AI what every part of their stack does. A rules engine checks each connection against sourced facts, and coding agents get the plan over MCP.

A tower-defense game rebuilt from a college team project. I replaced the engine after finding the movement, damage, and economy systems were each quietly broken, then grew it into a full game.

An autonomous robot that climbs vertical glass and draws on it, using suction adhesion, holonomic motion, and computer-vision docking. I led the electrical design.

An Arduino-driven pan-tilt scanner that sweeps an IR sensor across an object and reconstructs its shape in MATLAB: calibrated optics, clean serial pipeline, real results.