Projects

This is a course project for Text Mining and NLP at Ecole Polytechnique aimed at building a classification model that can differentiate between AI-generated and human-written text. Pre-trained language models such as BERT, RoBERTa, and GPT-2 were utilized to train multiple models, with the best model achieving a 90% accuracy after extensive testing. The project has implications in critical domains such as journalism, social media, and customer service.

This project aimed to quantify uncertainty in object detection models. To this end, I trained a Deep Ensemble of object detections models. Specifically, I proposed a new training protocol for training an ensemble of Faster R-CNN models using only one Region Proposal Network (RPN). In this training scheme, all you need is one RPN network and multiple Fast R-CNN prediction heads to create a robust ensemble network. This proposal produces results that are comparable to the traditional approach of fully training deep learning models end-to-end for ensembling.

This project tackles multiple object tracking which is a challenging problem in Computer Vision. To this end,I trained a robust pedestrain object model using Faster R-CNN object detection algorithms to localise pedestrains in video frame sequences using MOT17DET dataset. I then leveraged tracking-by-detection algorithms such as SORT and Tracktor++ for tracking people in MOT17 dataset.

This project is an implementation of a SLAM package for the Tello drone using the ORBSLAM2 algorithm. The ROS package achieved near real-time SLAM results and enabled the Tello drone to build a map of an unknown environment and to simultaneously determine its location within the map.

H-index Prediction was a class based kaggle challenge with over 120 groups in the Machine and Deep Learning Course at Ecole Polytechnique. The objective was to find the best NLP model, based on the MSE (Mean Squared Error), to predict the h-index of scientific authors in the domain of computer science.

This project exploit parallelism in a sequential image filtering program. This project aims to parallelize this program using parallel programming tools (OpenMP, MPI, and CUDA). This program reduced the computation time of applying Sobel and blur filters on gif images by 20%. The speed-up of image filtering programs is necessary as it is applicable in many sub-fields of image processing.

This project aimed to solve the non-convex optimization problem of finding an optimal path for a UAV to move from a base station to a target/inspection point while respecting the strict constraints of power, and data transmission rate. The problem was first transformed into a convex optimization problem and solved using classical optimization algorithms in CVXPY.

PlantDoc is a deep learning solution for diagnosing and recommending treatment options for African food crops. The model currently identifies some diseases in apples, maize, blueberry, potato, tomato, and cherries. This is a work in progress, and I plan to include in this project a recommendation system to suggest treatment options for the identified diseases. Also, I would build a mobile application to make this model easily accessible to farmers so that they can use this application on any smart device.

HomDoc is a telehealth application that enables people to have access to medical care outside of their physical geographical region. This application incorporates call, chat, and video functionalities that allows patients to communicate with medical doctors over the internet. The application has helped to provide improved access to quality healthcare in Ghana and Nigeria.

I trained an LSTM deep learning model to predict rainfall patterns for Kumasi. I built an efficient and smart irrigation system to conserve water using predictions from the deep learning model. I also created a mobile app to provide real-time farm data to the farmer.