About Me
Hi, I'm Armin Bagrat Stepanyan. If you've seen a different spelling of my name, this is the reason why.
I'm the founder of runspx.com and recurse.ml.
In my free time, I research energy-efficient learning algorithms. I do so by creating fast (<5 min) training benchmarks and iterating on them.
Here's one for LM architectures: github.com/cybertronai/wikitext.
And here's one for optimizers: github.com/cybertronai/caffeine.
Before, I was an NLP Research Engineer at an edtech startup, Boclips. There, I was the second member of the data science team. I built a video recommendation system, improved search, and led the development of a demo deployment pipeline.
I obtained an MPhil with distinction from the University of Cambridge by building the first GNN for predicting edges in DAGs. I used it to create a multilingual model for finding the minimum number of prerequisites required for a human to learn a topic. I was supervised by Ekaterina Kochmar and highly recommend her as a supervisor if you're interested in working at the intersection of NLP and education.
I was also an associate lecturer at Anglia Ruskin University, where I taught Fundamentals of Artificial Intelligence together with George Wilson.
I did my BSc at the University of Manchester. There, I mostly worked on side projects instead of going to class1. I did NLP research with André Freitas. We published it here: Identifying and Explaining Discriminative Attributes, (EMNLP), Hong Kong, 2019. I also led a team that created the best MERS prediction model for IARPA.
Fun fact: I've earned a US High School Diploma without ever setting foot on US soil.
And of course, you wouldn't be the first to point out that my middle name sounds an awful lot like a common RPG design problem.