Natural Language Processing (NLP) Developer in Berlin, Germany

A text annotation SaaS I built and run myself. LightTag includes interfaces for annotation, management of a distributed workforce, data quality assurance, and active learning.
I built LightTag because I needed it and turned it into a profitable business through a combination of ML and UX.

Ylabel is a POC project demonstrating full-text search powered annotation. The core of the project is storing an inverted index in the browsers IndexDB, which allows persistent full-text search without any infrastructure.

This project was an attempt to implement the "Generating Sentences from a Continuous Space" paper using Densenet as a variational encoder. Convolutional architectures are very effective with text and I did this project to see if they would make variational embeddings easier to train (they don't).

A long-form article aimed at practitioners discussing the merits and techniques of processing text with convolutional models as opposed to recurrent neural networks. Convolutions offer significant compute advantages because they are easily parallelizable thus shortening research cycles and lowering production compute costs and complexity.

A long-form guide I wrote as part of my work on LightTag. In the guide, I go over the typical lifecycle of a text annotation project as a pre-cursor to NLP and highlight best practices I've learned from my own experience and the experience of my customers.

Problem sets and presentations for a 2-day course I teach on NLP for practitioners. The course revolves around a single problem, converting German numbers into digits (Fünftausendvierhunderteinunddreißig =>5431) and focuses on pre-processing, regex, and tokenization before introducing sequential NLP models.

As a side project on a vacation, I implemented a few reinforcement learning algorithms attempting to optimally structure a portfolio of equities taking transaction costs into account.
After a foray into modern methods, I focused on a paper from the early '90s (Learning to Trade via Direct Reinforcement by Moody) that offers a much more domain focused approach to policy gradient algorithms.