Kyle Roth Email & Phone Number

Montreal, Qc, Ca

Montreal, Quebec, Canada

A project to augment large language models (LLMs) so they better understand procedural knowledge. Retrieval-augmented generation (RAG) is widely used to improve the factuality of LLMs' outputs, but can struggle to collect the right information to produce accurate procedures when that information is spread out and compositional in nature. We are developing a method called analogy-augmented generation (AAG) to address this, and plan to demonstrate this on a dataset of procedures extracted from… Show more A project to augment large language models (LLMs) so they better understand procedural knowledge. Retrieval-augmented generation (RAG) is widely used to improve the factuality of LLMs' outputs, but can struggle to collect the right information to produce accurate procedures when that information is spread out and compositional in nature. We are developing a method called analogy-augmented generation (AAG) to address this, and plan to demonstrate this on a dataset of procedures extracted from LangChain tutorials Show less

Provo, Utah Area

I built an online training service to manage parallel training of Kaldi models on sensitive live data. This involved implementing efficient parallelization and pipelining in the Go code, while also ensuring any single training sample could be deleted at any time minimal training interruption. I also implemented state-of-the-art learning rate algorithms for online training.

Provo, Utah Area

I worked on a project to simulate silicon photonic components using neural networks. The goal was to use neural networks as a drop-in replacement for expensive FDTD simulations when designing chip components.While working on this project, I also developed a tool called SLURM_gen to automatically generate and manage simulated datasets in a SLURM environment.I built a tool called SLURM_gen which helps a user manage datasets that can be generated in a high-performance computing… Show more I worked on a project to simulate silicon photonic components using neural networks. The goal was to use neural networks as a drop-in replacement for expensive FDTD simulations when designing chip components.While working on this project, I also developed a tool called SLURM_gen to automatically generate and manage simulated datasets in a SLURM environment.I built a tool called SLURM_gen which helps a user manage datasets that can be generated in a high-performance computing environment. Show less

Provo, Utah Area

I applied a (then) recent implementation of the variable-order conditional random field (VoCRF) algorithm to a morphological parsing task in Basque, achieving 71.3% accuracy. I worked with Dr. Lonsdale of the Linguistics and English Language department at BYU.

Greater Chicago Area

I wrote a fast C++ program to analyze how our competitors act in the Eurodollar market. It parses data from the exchange using the CME's SBE parser and the High Frequency FIX parser.I also built an interactive tool to display the actions of competitors in Brazilian currency futures using the Bokeh Python library. This is helping the company prepare to enter that market, as it works differently than the other markets they've entered before.

Provo, Utah Area

I improved the accuracy of a neural network ASR (automatic speech recognition) model from 76% to 94% using class-based (Thrax) language models. The model was specifically trained for autonomous drone recognition of air traffic control speech.I also wrote Go scripts to validate and generate spec files for a natural language understanding engine.

Helped college students understand calculus 1, 2, 3, differential equations, and linear algebra.

Doctor Of Philosophy - Phd, Computer Science, 3.7 Gpa

I started as a master's student in fall 2021, and in fall 2022 transitioned to a PhD position through accelerated admission. My master's.

Applied And Computational Mathematics, Applied Mathematics, 3.91

Activities and Societies: Data science club Applied and Computational Mathematics Emphasis (ACME) focuses on optimization, statistics.

Dual Credit Student