Sean Patrick Morris

Problem

Deep learning frameworks abstract away the mechanics of backpropagation and gradient flow, making it difficult to build genuine intuition for how neural networks learn.

Approach

• Implemented tensor operations, a computation graph, and forward/backward passes for linear and activation layers entirely in NumPy.
• Replicated PyTorch-style optimizer and neural net module abstractions to support gradient-based training end-to-end.

Result

A fully functional deep learning library capable of training multi-layer networks, demonstrating mastery of the mathematical foundations underlying modern ML frameworks.

Problem

Manual supplier category assignment was time-intensive and error-prone at scale, creating a bottleneck in procurement analytics workflows.

Approach

• Designed and deployed an enterprise agentic automation system on Palantir Foundry to classify supplier categories without human intervention.

Result

Reduced manual effort by 50%+ and integrated the system into live production workflows used across the organization.

Problem

Large-scale consumer review data is unstructured and noisy — extracting signal for sentiment and topic structure requires careful feature engineering and model selection.

Approach

• Applied TF-IDF embeddings with PCA, t-SNE, KMeans, DBSCAN, and hierarchical clustering to map the review landscape.
• Trained Random Forest, SVM, and logistic regression classifiers on binary sentiment labels.

Result

85% accuracy on sentiment prediction across a 100,000-review subset drawn from a corpus of 233 million records.

Problem

Training an RL agent on a real game requires bridging a hardware emulator, visual observations, and a reward signal derived from raw RAM — none of which are available out of the box.

Approach

• Wrapped the py-desmume Nintendo DS emulator in a custom OpenAI Gymnasium environment with grayscale 84×84 vision, 4-frame stacking, and RAM-extracted rewards for speed, lap progress, and collision penalties.
• Trained a Stable-Baselines3 DQN agent with a CNN policy on CUDA; built a live training viewer displaying metrics and action probabilities.

Result

Agent successfully completes a full lap on Yoshi Falls, demonstrating end-to-end RL from emulator integration through policy training to verifiable in-game behavior.

Problem

Global emissions datasets (OWID) are large and heterogeneous — extracting country-level trends and forward projections requires a reproducible, composable pipeline.

Approach

• Built a proper Python package (pyproject.toml) with modules for ingestion, preprocessing, scikit-learn forecasting, and multi-country comparison plots.
• Integrated CodeCarbon to track the tool's own emissions footprint per run, making the environmental cost of analysis explicit.

Result

A fully installable CLI tool with batch forecasting, grid visualizations, and a self-auditing carbon log — packaged for reuse across datasets.

Problem

Raw F1 telemetry data (speed, throttle, gear, braking) is rich but not easily comparable across drivers without an interactive visualization layer.

Approach

• Pulled 2024 race telemetry via the FastF1 API and stored it in a local DuckDB database for fast querying.
• Built a Streamlit app with animated matplotlib track maps, color-coded telemetry overlays, and play/pause/scrub controls.

Result

Live-deployed app at dsan5200group25.streamlit.app enabling lap-by-lap driver comparison across any 2024 race.

Problem

Demonstrating real-time object detection on arbitrary video input requires packaging inference, annotation, and a usable UI into a single deployable unit.

Approach

• Wrapped Ultralytics YOLOv8n frame-by-frame inference in a Gradio interface; annotated bounding boxes with class labels and confidence scores and converted output to animated GIF.

Result

Deployed on Hugging Face Spaces — detects 80 COCO object classes across uploaded video with no local setup required.

Problem

A fast-growing game studio needed to understand player behavior at scale to inform retention strategy, but lacked structured user segments.

Approach

• Applied KMeans + PCA to the full 770 TB Azure telemetry database to surface distinct, actionable user cohorts.
• Built churn prediction models using QDA and logistic regression to flag at-risk players early.

Result

Delivered user segments directly adopted for strategy optimization and a churn model with an 0.81 F1 score.