Ilya Kamen

A machine learning solution based on TensorFlow Object Detection that controls the correct mounting of parts in an assembly line for an automotive company. I developed the overall solution and led another engineer and an intern to complete certain milestones of this project.

The visual differences between correctly and incorrectly mounted parts were difficult for humans to perceive because they rely on small changes in the shape and shadow of the parts. The automated solution achieved a recall of 95% on incorrect mounts. In other words, 95% of the wrong mounts raised an alarm.

The solution is protected under an NDA, so the link points to a demo.

A solution that derives (noisy) labels for tail utterances. In a cross-functional team of engineers and scientists, I led the design and implementation of the solution.

Tail (rare) utterances can't be human-annotated because it's too costly, so this solution leveraged unsupervised and self-supervised machine learning algorithms. The project resulted in a double-digit improvement of the accuracy for tail utterances, and additional optimization had no adverse effect on average accuracy.

A command line interface (CLI) utility for developer productivity that I published as open source. Flynt transforms any type of string formatting and concatenation in Python into f-strings, the modern formatting method that is more performant and readable.