Michael Sheroubi

Mistral AI Robotics Hackathon - 1st Place

Using SO100 Arm with HuggingFace's LeRobot

This project showcases a mix of traditional robotics (optimal control) and AI Robotics (in this case we used Action Chunking Transformers). For the movement we went with a coded algorithm following an image processing pipeline. For detecting trash, we used a VLM with in-context learning. For the robotic arm policy, we used ACT with ~150 demonstrations. The robot was able to pick up trash and place it in a bin. The project was built in 36 hours and won first place at the Mistral AI Hackathon in Paris.

Comparing Spatial Reasoning in Discrete Environments

From Mistral Game Jam - Hackathon by Spotlight Paris (Entrepreneur First)

This study compares the performance of a Multimodal Model (Pixtral Large) and a Large Language Model (Mistral Large) in navigating identical game states. The Multimodal Model processes image based inputs, while the LLM relies on textual descriptions derived from the same images. The evaluation focuses on their decision-making, reasoning, and adaptability to provide insights into their respective strengths and limitations

Four Leaf Clover Detection

Computer Vision, Machine Learning
Using image processing to detect four leaf clovers

This project involves using computer vision techniques to detect four leaf clovers in images. The algorithm processes the image to identify and count the number of leaves, distinguishing four leaf clovers from the common three leaf variety. This project demonstrates the application of image processing and machine learning in a fun and practical way.

Roubi Analytics

Founder & CTO
Data & AI Research & Consulting

This company started as a way to help small businesses make data-driven decisions & build out their data infrastructure. With time I acquired bigger clients with multiple recurrent contracts working on AI projects.

After a personal shift in priorities, the company now focuses on supporting AI research and deployment in critical systems & infrastructure. I have since expanded the company to include a small team of researchers and engineers, with our ethos being "AI for Good".

Run-AI

AI, Computer Vision, Machine Learning
Using OpenCV to train a Convolutional Neural Network

This AI is designed to play the popular flash game "Run". I acted as sole/principal developer for this project. Using OpenCV's edge detection and a movemement vector as input, an image classification CNN, AlexNet, was fed hours of gameplay data to train the model. This project was meant to serve as an introduction into Machine Learning, but turned out to also be a great lesson in computer vision and image processing.

GovDB

Large Scale Web-Scraping, Vector Databases, Search Engine
Vector Database housing meeting minutes from municipal governments accross Canada

Designed to help contractors find and bid on government contracts. The database is updated routinely with meeting minutes from municipal governments across Canada. The search engine uses a vector database to find relevant documents based on user queries. Documents are split into chunks and converted into vector embeddings.

Text-Based Clustering

LLMs, Nature Language Processing
Use cosine similarity to create clusters from tokenized text

Often when working with large text datasets it is difficult to organize / categorize this data without pre-existing labels. By using the state of the art embeddings, we can find communities in the data by computing the cosine similarity between pairs.

Sudoku

Algorithms and Data Structures, Problem Solving, P vs NP
Failed attempts at solving sudoku in polynomial time

Periodically I to try and find an algorithm for a sudoku solver in polynomial time. One of my latest attempts involved converting a sudoku board into a system of equations and attempting to solve that. Other attempts included starting with a completed board and shuffling rows and columns (like a rubix cube) to match hints on starting board. I eventually hit a wall, but every idea is always very exciting. I obsessively chase each idea and enjoy each attempt.

Analog Clock Reader

Computer Vision, Problem Solving, Test Automation
Read time from an Analog Clock

The purpose of this project is to develop an algorithm for analog clock identification and to implement a program that would be able to read a given image or feed and return the value displayed. This is achieved broadly in three steps: find the clock, identify the centre, and highlight hands. The code utilizes Canny edge detection with Hough transform, and by calculating the angles between the hands to return the time. In the future, the code can be tweaked to read any kind of analog instrument, like gauge readers.”.

Benchmarking graph database performance

Databases, SQL/NoSQL, Query Optimization
Honours Thesis

Comparing the performance of Neo4j and SQL Server inside a social media application. The first phase involved data modelling, making application architecture decisions, and generating & importing data. The focus shifted towards designing queries to get information from the social network. The final phase was creating test procedures that collected perfomance data for each database.

Document-Based QnA (RAG)

Langchain, LLMs, Vector Databases
Retrieval Augmented Chat Agent powered by Llama2-7b and langchain

Just another RAG agent... built a few variations since this one.

Data Analysis: Video Game Sales

Data Analysis
Breakdown of video game sales from 1980 - 2020

Album Central

Web Development
Django Web App for reviewing albums

Nova-Graph

Databases, Graph Theory, Natural Language Processing
Graph Theory and NLP Research

A new approach towards building the "memory" of a virtual agent, NOVA. The aim is to create a knowledge-graph in high-dimensional space. Nodes would form relationships based on their position in this space. Their position in this space will change based on input from training texts. In short, each word would be defined by its neighbours inside this graph. The challenge is how to extract meaningful relationships and nodes from a piece of text (Big NLP Problem 1) and how to display query results in a meaningful output (NLP Problem 2).