Liam Frank

Experience

Junior Data Scientist

SAIM | Overland Park, KS

Janurary 2025 - Present

Work on a cross-functional team with Software, IoT and Database Engineers within the B2B SAAS space
Designed and deployed a transfer learning-based convolution neural network to label fuel facility assets in AutoCAD drawings, increasing previous model accuracy by over 20% to 98%, automating previously required manual data labelling
Developed an autoencoder model to detect anomalies in high frequency centrifugal pump vibration distributions, enabling predictive maintenance capabilities and early fault detection
Deployed an Azure Stream Analytics pipeline to over 50 commercial aviation fuel farms natiowide to monitor real-time sensor data for anomalies and quality issues, improving system reliability and detecting potential faulty sensors
Designed and implemented an automated hiring forecast model and reporting system that analyzed historical project data, discipline-specific utilization rates, and future projects to generate a strategic hiring plan and determine optimal staffing levels for each engineering discipline

Data Science Intern

SAIM | Overland Park, KS

May 2024 - August 2024

Developed an ensemble time series forecast comprised of an ARIMA, Holt's Exponential Smoothing, and definite integration for predicting remaining useful life of fuel filter vessels using differential pressure data
Researched and tested the use of unsupervised machine learning algorithms for anomaly detection for centrifugal fuel pumps
Worked alongside Mechanical Engineers aiding in developing and revising pump train sensor requirements to effectively supplement predictive maintenance initiatives of SAIM and Argus Consulting

Masters in Data Science and Analytics

Graceland University | Lamoni, IA

August 2023 - December 2025

Bachelors of Science in Data Science

Bachelors of Arts in Mathematics

Graceland University | Lamoni, IA

August 2021 - December 2024

Deep Convolutional Neural Network built utilizing transfer learning and the VGG16 architecture to classify satellite images of Military Aircraft.
Dataset included near 7,000 satellite images of 11 distinct aircraft types as well as satellite image of bareland
Employed image augmentation techniques at each training epoch to combat data size limitations and prevent overfitting
Accurary of 96% on testing data
Deployed model to a streamlit app hosted via a huggingface space for real-time user uploaded aircraft classification

Deep Feed-Forward Artifical Neural Network and XGBoost to classify default status of home-equity loans
Trained on just 6,000 observations and 10 features, the ANN had an accuracy of 80% and the XGBoost model had an accuracy of 93% when tested

Performed cleaning of raw text data through punctuation removal, case standardization, stopword removal, spell checking, lemmatization, and whitespace tokenization.
Tokenized words were transformed into numerical features using 1-gram text vectorization and the bag-of-words model. Resulting in a matrix of dimensions 50,422 x 20,818.
Principal Component Analysis was employed for dimensionality reduction. Shrinking the dataset from over 20,000 features to just 1,000 while retaining 44% of total variance.
Synthetic Minority Oversampling Technique was utilized to address class imbalances in training data.
Numerous supervised machine learning classification models were tested including Naive Bayes, Decision Tree, Random Forest, XGBoost, and Support Vector Machines with both linear and non-linear kernel functions.
Best performing models were the Random Forest of 200 trees and XGBoost model of 400 trees. Both models featured an accuracy of 97% and F1 score of 0.97.