Demetrio Rodriguez

A major manufacturer of foods and beverages saw a great shift in consumer purchase behavior toward online retail. Most of its eCommerce revenue comes from selling its products on a well-established platform.

However, the transition was somewhat unsteady: some products were declared out-of-stock and taken off the platform, creating massive revenue losses. Yet, only a small portion of those products was experiencing supply-chain shortages. For most, it was a combination of missed metrics like "delivery window," "weeks of cover," "past orders fill rate," etc. The eCommerce platform did not share the inner workings of its algorithms.

To facilitate weekly planning, I developed a machine learning model to forecast this out-of-stock behavior two weeks in advance. I combined the metrics reported by the eCommerce platform, internal supply-chain data, marketing planning calendar, and more. The problem was formulated as a time-series classification and solved using gradient boosting trees with inputs being various weekly aggregates of the last ten weeks in combination with known future static factors (e.g., holidays and promotions). I automated the output into a dashboard and delivered it to the stakeholders every Monday.

Traffic signs come in all shapes and forms. And very often, the most important part of a traffic sign is the text below it, especially if the text says when the sign is valid, e.g., "MON 6 PM-8 PM." Those texts are supposed to be roughly standardized and structured. Thus our development team approached the problem of converting the texts into strict rules by creating comprehensive regular expressions. This worked very well for a while but slowly became unmaintainable, so a scalable approach became necessary.

As our team already had a very comprehensive Regex-based parser, my suggestion was not to train an end-to-end system but a text simplifier. It is almost a machine translation task: all of "MON," "MND," "Mondays" would become "Monday," "Noon-3 PM" would be translated as "12 PM-3 PM," "No Littering!" would be ignored.

For this problem, I trained a state-of-the-art (at the time) NLP model—a bidirectional LSTM with attention. After just two months of development, it achieved a reasonable accuracy (92%) and was suitable for a human-in-the-loop deployment. Additionally, we requested a research grant to scale the solution further.

Oftentimes machine learning solutions focus on predicting one number. Specifically, this is not always very useful in wholesales, as the actual daily sales can vary quite a bit. Such variation when not addressed will result either in overfilled storage or empty shelfs. In order to effectively perform capacity planning, a manager should know a range of outcomes that could happen with some degree of certainty.

In order to address this challenge, I've trained a mixture density neural network. The output layer in such architecture are parameters of a mixture of distributions (gamma in this case) and a parametric inverse maximum likelihood is used as a loss for training. This allows capturing multi-modal conditional distributions or a broad range of right-skewed distributions. As the data came from various stores across different geographic regions and exhibited strong trend shifts, it was first de-trended, then standardized before being modeled by the mixture density network.

Predicting on-street parking space occupancy is an extremely challenging problem. Mainly as there are no reliable sources of the ground truth.

Our solution was to use satellite imagery as a scalable approach to assess the parking situation in multiple cities across the globe at once. The main challenge is not to detect cars on the satellite images, which is just an object-detection problem (a very nasty one, however). It's about putting together a multi-stage pipeline that uses machine learning, heuristic rules, and legal restrictions to output how many free parking spots are there on a street.

The blog article was written by me and explains our approach in great detail.

A well-known car-sharing company approached us with a problem: some of their cars were being taken within minutes after being parked and some stand unused for days. They already knew it had to do with the geography of the city, population density, major transport hubs, time of day, and so on.

So, we offered to build a machine learning model that would take all of these influencing factors into account and tell where and when cars will be in high demand to initiate relocating cars from low-demand areas.

Under my supervision and mentoring, a junior data scientist on their first project and I delivered a successful MVP. We found a suitable deployment strategy through my cooperation with the project team and launched the first version of the product four months after getting the initial dataset.

Consequently, I took over scaling the solution to multiple cities, adjusting its real-time efficiency, and adding multiple features based on the client's requests and model performance.

Eventually, this has become the most successful product of the startup; they then rebranded and now offer it as their only service.