Stephen Sommer

Pen 1 · Pen 2 · Pen 3
50 µV ⌷ 1 s ⌷ Fp1 · Cz · O2

Software developer → machine learning scientist

Machine learning across signals, images, and decisions

I'm a software developer finishing an MS in Machine Learning at Georgia Tech, working toward a career as a machine learning scientist. My project work spans security classification, medical image segmentation, and reinforcement learning. By day I write Python and SQL pipelines on Databricks that move data into a warehouse and lake. Completed work below has a repo; work still in progress is marked as such.

Résumé (PDF) GitHub Email

Selected work

Three of them

Security ML · poster, Winona RCA 2024

Decision trees for remote-to-local intrusion detection

Compared decision-tree detection of R2L attacks trained on KDD Cup 1999, on NSL-KDD, and on the two combined. Merging the datasets raised accuracy to 99.93% with a 91% detection rate and a 0.08% false-alarm rate, beating either dataset alone. Presented at the Winona State Research and Creative Achievement Celebration.

Accuracy 99.93% Detection 91% False alarm 0.08% Data KDD99 + NSL-KDD

Medical imaging · research assistantship

3D prostate MRI segmentation with nnU-Net v2

Trained and evaluated 3D segmentation models on Task05 of the Medical Segmentation Decathlon as a CS research assistant at Winona State. Work covered preprocessing multi-modal MRI volumes, configuring nnU-Net v2 training, and reviewing predicted segmentations against the reference annotations in MITK.

Data MSD Task05 Model nnU-Net v2 Task 3D segmentation Role Research assistant

EEG decoding · deep learning · planned, fall 2026

Liminal EEG — predicting experience intensity from brain recordings

Planned research project analyzing 34-subject EEG recordings with MNE and PyTorch: ICA artifact removal, epoching, and Welch power-spectral-density features, benchmarking an EEGNet convolutional model and a temporal Transformer against a published SVM baseline under leave-one-subject-out cross-validation. Goal is a written, published paper.

Status Not yet started Data 34 subjects Models EEGNet · Transformer Goal Published paper

Also on GitHub: a statistical analysis utility and a modular voice-assistant sandbox → github.com/ssommera

Instruments

What I work in
Modeling
  • PyTorch
  • TensorFlow / Keras
  • scikit-learn
  • Reinforcement learning
  • Transformers & NLP
Signal & imaging
  • MNE-Python
  • ICA artifact removal
  • Welch PSD / band power
  • nnU-Net v2
  • Leave-one-subject-out CV
Data engineering
  • Databricks
  • Python ETL pipelines
  • Oracle PL/SQL
  • Warehouse & lake modeling
  • Power BI
Platform
  • Python · Java · SQL
  • Terraform
  • Azure AI Foundry
  • Git
  • Secure development (OWASP)

Record

Education & roles
In progress
MS Computer Science, Machine Learning Georgia Institute of Technology. Coursework: Artificial Intelligence, Machine Learning, Reinforcement Learning, Data & Visual Analytics, Game AI.
2023 —
Software Developer, Fastenal Python and Oracle PL/SQL ETL on Databricks feeding the data warehouse and data lake; data cleaning, curation, and ad-hoc analysis. Member of the Best Practices and AI Dev Efficiency groups.
2024
CS Research Assistant, Winona State 3D prostate MRI segmentation with nnU-Net v2, and decision-tree intrusion detection presented at the university research celebration.
2022 — 2024
Software Developer, Winona State Built and maintained the university and CS department sites, configured course VMs, supported lab technology, and tutored students.
2024
BS Computer Science, Winona State 3.86 GPA. Minor in Psychology (neuroscience and cognitive science). CS Club president. Preceded by an AS in Computer Science, RCTC, 4.0 GPA, completed while working full time.

What I want to work on

Research
Neural decoding

Learning representations from EEG and other biosignals that hold up across subjects, not just across epochs from the same recording.

Better algorithms, not bigger runs

Sample efficiency and stability in reinforcement learning, and understanding why a method works rather than only that it scored well.

Machine minds

What it would take to make claims about machine awareness testable, and what alignment research owes to philosophy of mind.

Get in touch

Minnesota · open to remote
Emailssommer.ai LinkedInStephen Sommer GitHubssommera Stack OverflowProfile