Achraf Bennis

AI Market Research Agent: From Data Collection (AI Interview) to Insight Delivery and Reporting

Backed by Skydeck US (1% acceptance rate)

Using Machine Learning, Deep Learning and Data Science techniques as well as computer vision, we implemented the following pipeline: ○ Reading and pre-processing of microscopic images containing cells (cancerous and normal) using the OpenSlide library in Python. ○ Cutting the images into sub-images (patches) and selecting those that are likely to have cancerous cells, in order to reduce the computation time. The selection is a procedure based on different mathematical filters such as the Laplacian filter (implemented in OpenCV Python), scores formulated using the expertise of hematologists and oncologists. ○ Segmentation and extraction of cell images from selected patches: using Deep Learning methods such as Faster R-CNN, SPP-NET, etc. of the classification (provided by WHO 2022) of the different types of cancer cells responsible for leukemia: K-means, Affinity Propagation Clustering (Scikit-learn). ○ Implementation of an algorithm that predicts the genetic mutation of cancer cells in order to predict a treatment adapted to the patient. The method will be inspired by survival analysis based on Deep Learning (DPWTE, Cox-PH, MLP, etc.) or Machine Learning (Random Forest, SVC, XgBoost, etc.) methods. All models will be implemented in Python (PyTorch and Scikit-learn). ○ Implementation of a Python library automating all pipelining with Python software standards (Pep 8).

Skills: Computing · Responsible AI · Software Project Management · Diffusion · Generative AI · Generative AI Tools · IT Strategy · AI Software Development · Programming

The goal of this project was to analyze the movement of an operator working in a sorting center through digitized video recordings: ○ Transformation of the time series (describing the movement of the joints of the different body members) for data smoothing: application of mathematical filters such as ’Butterworth filter’, FFT, etc. using the Python package Scipy. ○ Analysis of the time series after transformation: calculation of the speed and acceleration of the different joints and comparison with the video to validate the applied transformation. ○ Implementation of a K-means-based clustering method using the DTW (Dynamic Time Warping) metric to compare two body movement time sequences. ○ Writing reports explaining the methods implemented and the results obtained.

The aim of the project (in collaboration with ADEME and ENEDIS) ’SmartOccitania’ is to implement algorithms based on Machine Learning to automate the detection of faults in the electricity networks of the region ’Occitanie’: ○ Pre-processing of network and fault history data. ○ Implementation of classical Machine Learning models (Random Forest, XgBoost, SVM, Logistic Regression) for network classification (at risk: yes/no). ○ Implementation of DeepWeiSurv and DPWTE, Deep Learning-based approaches (designed during my Ph.D. thesis) inspired by survival analysis (statistics) using PyTorch. ○ Comparison of different approaches using scores: Accuracy, F1-score, etc. ○ Comparison (in predictive performance) between the ENEDIS expert method and the validated Machine Learning model whose performance is significantly superior. ○ Drafting of reports describing the pipelines implemented for this project.

○ Pre-processing of the ’xslx’ file containing the claim (text) reports of different languages: stemming, removing stop words, and applying the TF-IDF statistical function to represent the text data when implementing the models. ○ Implementation (using the R language) of different Machine Learning methods (SVM, RandomForest, XgBoost, Logistic Regression) to predict the probability that a claim is fair. ○ Comparison and validation of implemented models using metrics such as F1-score, AUC, Recall, Precision, etc. ○ Implementation of an application using the Shiny package of the R language which hosts the validated and production-ready models.

○Pre-processing of audio files (’.wav’ format for example) from public databases SIWIS and ESTER: digitization (in frequency and amplitude), transformation (using mathematical filters such as a log-filter bank). ○ Extraction of features from the transformed audio inputs: mean and variances of amplitudes/frequencies, and other statistics using the Python Scipy library and associated transformations using log-fbank for example. ○ Implementation of Machine Learning (Random Forest, XgBoost, and SVM Classifier) and Deep Learning (CNN) methods for the detection of pronounced words with accentuation while taking into account the context. For this, we trained our models with SIWIS audio files and/or their features as well as annotation files. Libraries used: Keras/Tensorflow and Scikit-Learn. ○ Comparison and validation (k-fold cross-validation) of models using scores such as F1-score, Sensitivity (Recall), Specificity, and Accuracy

Implementation of a Deep Learning method (MLP) for solving partial differential equations using the Scikit-learn and Keras libraries.