The objective is to identify road defects on airports runways. As explained by our coaches at Colas for this project, there are different types of road defects (faïençage, fissure, etc...).
I was in charge of the data engineering on this project, which consisted in requesting the orthophotos and delimiting them on the airports. There was quite a lot of geospatial thinking involved, and I got familiar with GetMap requests.
Then, we implemented one of HuggingFace's transformers for Computer Vision tasks.
Starting with a dataset containing approximately 500 hospital admission notes, we had to label data and build a NER (also called token-classification) model.
The goal is to augment the dataset, and train a model to extract as many of the following indicators for posology : the drug, the form, the dosage, the duration...
We relied on HuggingFace pre-trained LLM such as BertForTokenClassification and Camembert
I was in charge of building the pipeline and training the models.
QuantumBlack's challenge was focused on two computer vision tasks : image classification, and then segmentation. We also built a Streamlit app that encapsulates our CV model, and helps to visualize our business recommendations. I was in charge of building the Streamlit app, and coming up with business recommendations.
It was a project on time series. For each server's CPU and memory, we had 500 data points corresponding to 500 days. A data point is actually three values: max_value, avg_value and min_value. The objective is to optimize the CPU and memory settings (number of core, mem size) to avoid both saturation and idleness.
I was in charge of data processing, more particularly analysing and validating periodicity within the time series. So that we could feed only relevant signals to our Prophet time-series model.
We implemented an algo called CCF-iterate available here: https://www.cse.unr.edu/~hkardes/pdfs/ccf.pdf We used Scala and Python, and wrote programs using RDD (Resilient Distributed Datasets) or Spark DataFrame and tested both methods on increasingly big graphs. I was in charge of writing Python code.