Mihai – Python, AWS, SQL

An energy startup looking to change the way the electricity and gas markets operate. The project includes two main goals:

• predict the future wind/solar production for different producers;
• predict the future electricity/gas demand on the market for different distributers in order to optimize the efficacy of the market and reduce costs when energy needs to be sold/bought on the market in a short notice period. Mihai started this project from almost zero by first building the ML algorithms and then slowly building a team to take care of the ETL, the platform, and the deployment of the algorithms in production.

• built several algorithms for predicting wind/solar production and timeseries algorithms for predicting future demand in the first month;
• used the wind/solar production algorithms successfully for 20 clients in pilot projects;
• the wind/solar production algorithms placed top 3 in a live competition together with big players already existing in the energy prediction market;
• the demand algorithms ran successfully for 5 big companies in pilot projects, all of which decided to move towards implementation in live production;
• implemented all the training pipelines in AWS lambda for parallelizing and speeding up the training;
• led a team of 2 DS and 2 ML Engineers for incremental improvements of the original algorithms and a stable production environment;
• could onboard anybody on the codebase in less than a week because everything was very clearly written and organized;
• was responsible for the ML part of due diligence for the first round of funding in which we gathered 2 million dollars.

The biggest online advertising company in the DACH region, working both with the supply and demand side of the online advertising space. The project Mihai worked on was building a real-time prediction algorithm that could forecast the number of ads a certain campaign could realize under the time and budget constraints a campaign manager would set up.

• worked in a team of 2 to set up a 4 step auction win prediction algorithm in production in 3 months;
• improved auction win accuracy up to 95% using boosting methods;
• deployed algorithm in dockerized fashion via AWS ECR.

An M&A facilitator for small and medium businesses in the DACH region. The goal of the project was to make the life of the consultants easier by building an M&A recommendation engine based on modern NLP techniques.

• worked in a team of 2 to design a POC making use of new NLP embedding techniques (BERT, ALBERTA);
• performed lots of data cleaning and exploration for unstructured text data obtained via scrapping;
• helped in writing and obtaining a 1 million euro governmental grant for improving the results of the recommendation engine.