AI Engineer
Resume:
JAMES BEDICHEK
CONTACT PROFILE
ad7a0m@r.postjobfree.com
www.linkedin.com/in/james-
bedichek-0a83a4213/
Passionate and creative AI enthusiast with an easy going and positive attitude. Looking for an opportunity to apply and expand my understanding. Particularly excited about deep learning theory and research.
Solid PyTorch skills with experience building complex models from scratch.
Solid Python skills with experience using FastAPI, NiceGUI, Pandas, NumPy, OpenCV, and Pillow.
Experience collaboratively designing AI experiments in a professional context.
Currently enrolled in Foothill College, expecting to obtain AA in Computer Science in the Fall of 2024.
SKILLS EXPERIENCE
Python
PyTorch
Linear Algebra
Statistics
Diffusion Models
Computer Vision
Transformers
AI/Software Intern @ Mihira AI
Jan 2024 - Jun 2024
- Designed and implemented a Bayesian hyperparameter optimization program for doing PEFT on Stable Diffusion in PyTorch
- Designed and implemented a generative Diffusion model from scratch in PyTorch, trained on the MNIST digit dataset.
- Wrote a program to classify human poses in training data based on Open-Pose key point detection.
- Contributed to a Fast API web application program in Python. Implemented a front end GUI and various
EDUCATION
University of Colorado
Boulder
2020-2021
Computer Science
Cornell Machine Learning
Certificate
2021-2022
Foothill College
2021-Present
Computer Science
Relevant Coursework:
Multivariate Calculus
Linear Algebra
Discrete Math
Python Data Structures/
Algorithms
Computer Architecture
Python. Implemented a front end GUI and various
endpoints for automated LoRA training.
- Contributed to the design and execution of experiments on LoRA training of Stable Diffusion
Deep Neural Network Stock Price Predictor (hobby project) Feb 2024 - Present
Wrote Python code to scrape hundreds of gigabytes of stock, language, and market data from Yahoo Finance, then feed it into a deep neural network. It’s current iteration predicts short term price movements as a probability distribution, taking daily stock movements, market data, sector, and language information of a company daily for a 3 year period to create a sequence, and then feed that sequence to a vanilla transformer architecture, using greedy layer-wise training. Earlier iterations used deep-Q Reinforcement Learning. It is trained on over 3 million data points from 2600 companies, using one 4090 GPU. The model integrates relatively new developments in optimization and training methodology, like EL2N dataset pruning and the Lion optimizer. It’s current iteration shows high accuracy, but unstable performance on test sets (averaging ~20-30% profit over 3 months, depending on simulation parameters). The whole project is currently over 2000 lines of code, with an additional 1000 lines of code from earlier design iterations. Some features in the project which demonstrate my
understanding of statistics and math related to ML are:
- Using the entropy of probability distributions to estimate the uncertainty of the model’s prediction.
- Different types of data normalization like unit cube and gaussian.
- PCA to compress market data.
- Gaussian noising of data for regularization in training and averaging in inference.
- Using Sinusoidal embeddings to represent categorical information, like the sector and industry of the company.
(Got this idea from the time-step embedding mechanism in Stable Diffusion).
Project GitHub:
https://github.com/JBedichek/Stock-Prediction
Contact this candidate