I build machine learning models and study how their design changes what they get right and what they cost to run.
Right now: research assistant to Prof. Shuai Xu, robotics software for CWRUbotix, and teaching assistant for Introduction to AI.
A toy router with fixed random weights, so you can watch the mechanism. In the real model the router learns its weights in training.
A mixture-of-experts language model I wrote and trained from scratch in PyTorch. It is a 60.6M-parameter decoder-only transformer. For every word, a learned router picks two expert networks out of several, so only part of the model runs for each word.
It uses rotary position embeddings and generates text with a key-value cache. I trained it on 20M tokens of TinyStories.
Against a dense model trained on the same 20M tokens, the sparse model reached 6.9% lower validation perplexity. Lower is better: it was less surprised by stories it had not seen.
| Model | Perplexity |
|---|---|
| Sparse, top-2 experts | 10.03 |
| Dense baseline | 10.77 |
A model can train and still be wrong in quiet ways, so I wrote tests for the parts that fail silently:
Ongoing research
With Prof. Shuai Xu I study how to prune neural networks: removing weights so a model gets smaller and cheaper to run without giving up much accuracy. It is the same question as Story MoE from the other side. Instead of running only part of a large model, make the model itself smaller.
The work is in progress, so there are no results to show yet.
CWRUbotix, MATE ROV
CWRUbotix builds an underwater robot for the MATE ROV competition. I build real-time object detection pipelines with ROS and FLIR cameras.
I set up a Raspberry Pi Zero 2 W to stream video for a USB photosphere imaging system. During the team's bootcamp I also wrote ROS nodes in Python to drive a simulated robot.
Python, data analysis
A terminal program for comparing homes in a housing dataset. You choose the filters that matter to you. It ranks the matches with a weighted score and draws graphs to compare them: price distribution, prices by neighborhood, and area against price.
JavaScript game
A pet game where your choices change how it plays. It has a shop with its own money system and an inventory that the game keeps track of.
I’m working toward a BS in Computer Science and an MS in Machine Learning and Artificial Intelligence at Case Western Reserve, with minors in Business Management and Mathematics. I finish in May 2028.
I like knowing how things work, all the way down. That is why I wrote Story MoE from the attention layer up instead of loading a pretrained model, and why I wrote tests for the parts that could be quietly wrong.
The question I keep coming back to is cost: how a model’s design changes both what it predicts well and what it takes to run. Sparse experts and pruning are two answers to it.
Outside research, I write vision code for an underwater robot and teach introductory AI.
I like to see a project through, from the first version of the code to testing it and checking whether the results hold up.
Fotih Turopov
Cleveland, Ohio
Newest first, within each group.
Python, PyTorch, TensorFlow, ROS, Java, JavaScript, SQL, Git, Ubuntu
If you have an opportunity or a question about my work, I’d like to hear it. Email works best: fxt123@case.edu