Dragos Tudor

One of the projects that I've worked on is calculating the probability of a warehouse running out of stock given (Xt = Xt-1 + e^Wt, Xt) accumulated orders at a given time, and Wt is a normally distributed random variable with mean 0 and standard deviation = 1.5. Orders at t = 0 is X0 = 10. The warehouse has 100 units in stock. To do: estimate the probability of running out of stock at t=10.

I trained a CNN (convolutional neural network) in TensorFlow to recognize houses from satellite imagery. The aim was to re-run the model on an image post-earthquake for identifying collapsed units. 97%+ accuracy.

I built a model for visualizing clusters of road accidents across the UK. I used KDE and XGB for visualizing and modeling road accidents. Thousands of various accidents were analyzed in the process.

This is a project where the music adapts to your emotions with data extracted from your own webcam.

I designed a smart notification management system. It is an adaptive notifications system that aims to group together multiple notifications based on the users' circadian cycle, popularity, and behavior.

I boosted a UK-based industrial retail client's revenues by 11% by recommending opportunities to upsell.

This project consists of two main parts:
1. London housing market price predictions—stacked learners and seasonal ARIMA-based models.
2. Forecasted the error of Zillow's internal model better than 93% of other submitted models; used stacked models in Python.

I used RCNN/DCNN and CRF on 50,000+ samples (ISIC, scraped and generated imagery) for identifying over 1,000 skin condition subtypes from HD images.

The project's goal was to unify and structure the existing knowledge about dietary preferences, allergens, intolerances, and their interactions. Given the distribution of various allergens, both IgE and non-IgE (delayed response), the goal was to identify the likelihood of one of them to be present in a specific ingredient and, therefore, to pose a threat to the user of an app.

Some users might be allergic to peanuts, which might not be an issue if the dish contains Brazilian nuts. Similarly, a user might be intolerant to peanuts, but not if the amount is small in a given dish.

For the two types of users, the perceived risk can differ. In the first case, the user perceives Brazil nuts as dangerous, while their real risk is low (restaurant might also process groundnuts), and in the second case, the perceived risk is medium, but the user can decide if it’s acceptable. All of these allergen - ingredient ’risk’ relationships are approved by an expert and categorized.

Hospitals' DICOM imagery is curated, securely stored, anonymized, calibrated, and automatically annotated by using custom computer vision algorithms, at scale.

Access to anonymized imagery is offered on-demand, to verified research departments, startups, and other partners, via virtual machines (VMs) hosted in a private cloud with strict data management and exfiltration prevention protocols.