Rei Meguro

I enjoy the art of over-engineering complex, fast things

About Me

I'm an ambitious Computer Science student who loves to automate things. I work closely with all kinds of AI (ML, CV, NLP), as well as full-stack. I am also the author of Japan' strongest chess engine, Raphael.

My Recent Projects

Click on the cards to visit project links

Raphael

Raphael is a chess engine I wrote in C++ and it is currently the strongest chess engine in Japan.

It uses alpha-beta search with various enhancements, and a quantized neural network running on the CPU using SIMD intrinsics, trained on tens of millions of self-play games.

Onyx

Onyx is an open source AI platform which automatically ingests documents from connected sources and leverages deep research agents to answer questions or complete complex tasks.

At Onyx, I helped develop the deep research agent and the knowledge graph indexing & inference pipeline, eliminating OOMs, improving fault-tolerance, and tripling indexing speeds.

AdaPhish

AdaPhish is an adaptive phishing detection and submission platform which learns attack patterns over time using a vector store and RAG, built using Docker, FastAPI, Next.js, and Nginx.

It was published in the 4th IEEE ICAIC Conference and won the Best Paper Award.

Kineval

Kineval is a Python-based interactive learning tool for learning FK, IK, collision detection, and RRT motion planning.

It was developed to be used for EECS 367 at the University of Michigan, and it's written in a manner that allows students to implement these features on their own without worrying about other aspects such as rendering.

Folio

Folio is a startup which connects freelancers to venture-backed startups.

I developed the core email notification system and helped develop data migration pipelines to transition from their old platform, working with Django, Next.js, Docker, and Git.

ConverseCart Search Engine

Conversecart is a startup which helps retail stores greatly boost their product click through rates by providing them with a smart AI search engine.

I helped develop the early versions of their search engine by leveraging word embeddings, bag-of-word models, and cross encoders to enable fast, personalized, natural language queries.

RTD Toolkit

RTD, developed at the University of Michigan, allows provably safe motion planning accounting for uncertainties in measurements using polynomial zonotopes.

I worked with the ROAHM Lab to port their implemention of RTD from MATLAB into a generalizable library in Python so that anybody could implement RTD, whilst also introducing performance optimizations.

Mandelbrot Fractal Explorer

This is am interactive Mandelbrot Fractal Explorer I wrote in C++ using SFML.

It uses threadpools to exploit the embarrassingly parallel nature of the problem, allowing 1080p renders and zooms of up to 10^-16.

Sudoku Solver

This is a simple Sudoku Solver using a backtracking algorithm for the solver, a CNN for the digit recognition, and pyautogui for the automatic number entry.

Nerd Wrecker: Connect 4

This project is a fully automatic, multiplayer Connect 4 player.

It uses the alpha-beta algorithm paired with bitboards to efficiently find the optimal moves from the current position. It also leverages pyautogui to read the game states and enter commands into the GUI.

LC2K CPU Simulator

LC2K (Little Computer 2000) is a minimalistic ISA with 8 instructions, used throughout UMich's EECS 370 Computer Organization class.

This project implements LC2K along with a memory-mapped display within Excel, using self-referencing cells to simulate the clock, registers, and memories.

Plaite

Plaite is a nutrition tracking app which uses YOLOv8 for real time food detection, developed during the 24h MHacks 2024 Hackathon. It also comes with an LLM food recommendation assistant.

The project won the Best Use of Intel AI Award.