I build machine learning systems that have to survive contact with production.
I'm a senior at the University of Mississippi studying Computer Science with a Data Science emphasis and a Math minor, graduating May 2027.
My longer-term interest is embodied AI and robot learning, which is to say the version of this problem where the model has to act, not just predict. Everything I build keeps circling the same question, which is what breaks when a model leaves the notebook.
Finding the clauses that actually carry risk in commercial contracts, so a legal reviewer knows which document to open first instead of reading all of them.
- Designed the offset-preserving chunking strategy the team's pipeline runs on — 510-token sliding windows with 128-token overlap — keeping character offsets aligned to the source text so every clause prediction stays auditable, with 97.8% of annotated clauses intact inside a single chunk
- Disproved the team's leading hypothesis for classifier misses (clauses split across chunk boundaries: correlation +0.004, only 2% of spans split) and isolated positive training volume as the real driver (+0.715), redirecting the team away from a costly chunking rewrite
- Built 5-fold cross-validated threshold tuning split by contract, scoring every training contract with a model that never saw it — lifting held-out macro-F1 from 0.717 to 0.750 with no retraining, and the weakest clause category 5x (0.091 → 0.500)
- Showed per-category thresholds overfit on low-support categories (as few as 27 positive contracts) and shipped a single global cutoff instead
Python · PyTorch · Hugging Face Transformers · DistilRoBERTa · Pandas · scikit-learn
Five-person team project on the CUAD corpus — the work above is my own commits. Repo →
Turning raw malicious JavaScript into named attacker "families" so analysts can chase infrastructure instead of individual scripts.
- Shipped a production graph-ETL pipeline backfilling six months of threat data into a live Neo4j store through idempotent, rollback-tagged loads, automated as a nightly self-healing cron job
- Built a 7-signal attacker-profiling engine clustering malicious JS via SimHash AST similarity, perceptual screenshot hashing, and shared TLS cert / IP / ASN / registrant infrastructure, using GDS Leiden community detection
- Diagnosed and fixed a 52K-script over-merge caused by benign library bridges between unrelated families, cutting max family size to 476 across 334 clean clusters
- Extracted a CNN-LSTM-Attention fault predictor (94.7% accuracy) out of a legacy Django monolith into a clean FastAPI service on a 16-container microservices platform, with zero-downtime deploys
Python · Neo4j / Cypher · PyTorch · FastAPI · Docker · Nginx · Gunicorn
Proprietary work. Happy to walk through the architecture and the debugging story.
| Project | What It Does | Tech |
|---|---|---|
| Census Income Classification 📌 | End-to-end ML pipeline on 32K+ census records. Tuned on F1 rather than accuracy to handle class imbalance, found a 3-layer Keras net beat logistic regression by only 0.02 F1 (0.70 vs 0.68), too small to justify the added complexity, so shipped the simpler interpretable model. Audited coefficients for fairness, surfaced a discrimination risk, and recommended against lending deployment. | Python Scikit-learn Keras Pandas |
| Streaming Evaluation Protocol for CSLR 🔬 | A standardized evaluation protocol for streaming sign language recognition, adapting latency and stability metrics from streaming ASR and simultaneous MT to fill a gap left by offline-only benchmarks. Includes a fault-tolerant landmark extraction pipeline over PHOENIX-2014T. | Python MediaPipe Multiprocessing |
| Real-Time ASL Translation 🤟 | Live ASL recognition from webcam input, built for interactive frame rates on consumer hardware rather than batch inference. | Python TensorFlow OpenCV MediaPipe |
| Multi-Modal RAG Chatbot 🧠 | Gemini-powered agent with persistent long-term memory: a lightweight secondary model emits schema-enforced JSON add/update/delete operations over SQLite, classifying entries as semantic, procedural, or episodic and resolving contradictions in place rather than duplicating records. RAG over a document store and the memory layer, with multimodal image and document input. | Python Gemini SQLite Streamlit Docker |
| Hyntor 📚 | Responsible-AI study platform for web and mobile: an AI study partner with shared per-course memory, built so students get help that is scoped to their course rather than open-ended answers. | TypeScript React AI |
| CS2 Esports Statistics Tracker 📊 | Normalized relational schema for pro Counter-Strike teams, players, tournaments, and match performance, designed around the queries analysts actually run. | SQL Database Design |
Also on my profile: an e-commerce storefront in TypeScript/React and a desktop note-taking app in Python.
💻 Languages
🧠 ML & Computer Vision
☁️ Infrastructure & Data
Open to conversations about embodied AI, robot learning, and anything that puts a model on real hardware.


