ML engineer: computer vision, LLM fine-tuning, and MLOps. M.S. Artificial Intelligence (4.0), Kennesaw State University. Atlanta, GA, open to remote or relocation. U.S. citizen, eligible for a security clearance.

I train and fine-tune vision and language models, and I deploy them (Kubernetes, CI/CD, monitoring), usually under compute constraints that look a lot like edge and on-prem environments.

📫 LinkedIn · drewpatrick25@gmail.com

🚢 Enterprise Kubernetes MLOps Pipeline, team capstone (5 people) Real-time computer vision inference on Kubernetes as three containerized microservices (image intake, YOLOv8 detection, live dashboard). ArgoCD GitOps + Helm with self-healing and zero-downtime rollout, gated by GitHub Actions CI. DevSecOps controls via Trivy image scanning and sealed-secrets. Diagnosed OOM kills with Prometheus/Grafana and cut intake memory from ~2 GB to ~30 MB by caching pre-encoded JPEG frames instead of raw pixel arrays. Trained YOLOv8n-seg on an A100, then optimized it for CPU-only inference to fit constrained cluster resources. Five-person team; the model training, CPU-inference optimization, and the memory diagnosis were my workstreams. Kubernetes Docker ArgoCD Helm YOLOv8 Prometheus Grafana Trivy

⎈ Kubernetes Engineering Labs, solo 165 manifests, transcripts, and write-ups across twelve weeks: scheduling and HPA, probes and rollbacks, Helm, GitOps with Argo CD, Trivy image scanning, RBAC, default-deny NetworkPolicies, Pod Security Standards, and a Grafana-observed incident simulation. Kubernetes Helm Argo CD Trivy kind

🧩 Gameboard Genesis, LLaMA-2 + LoRA for structured rulebook generation OCR → Markdown corpus pipeline over 651 real board game rulebooks, 4-bit LoRA fine-tuning of a 13B model on a single A100, beam-search decoding. PyTorch Transformers PEFT/LoRA bitsandbytes Tesseract W&B

🔍 STAR-CAST, auditable unsupervised fraud detection Extended Isolation Forest with rule-card explanations, evaluated under a leakage-free time-aware protocol with validation-frozen thresholds and block-bootstrap CIs. Includes the finding that density baselines outperformed EIF on PCA-style features, reported rather than buried. scikit-learn PyOD Extended Isolation Forest pandas

⚡ BERT++, BERT-scale pretraining on constrained hardware Encoder pretraining with FSDP parameter sharding, gradient checkpointing, AMP, and FlashAttention over streamed corpora, with W&B tracking. PaLM-style parallel blocks, SwiGLU FFNs, XPos rotary embeddings. PyTorch FSDP Transformers FlashAttention CUDA

Graduate AI Research Assistant, Kennesaw State University (Jan-Jun 2025) Processed 1,000+ PET/MRI scans to compute standardized uptake value ratios (SUVRs) and built a cross-modal CNN fusing PET and MRI inputs for Alzheimer's classification, working with a multidisciplinary team on real, noisy clinical data.

Δ-Learning HSE Band Gaps from Low-Cost DFT, materials informatics Gradient Boosting / XGBoost surrogate predicting expensive hybrid-functional band gaps from cheap DFT plus Magpie composition descriptors (JARVIS-DFT). Composition-grouped splits to prevent leakage; MAE ≈ 0.20 eV vs. ≈ 1.07 eV for the uncorrected fast baseline.

Closing the Spatial Gap in Vision-Language Models (in progress) Injecting depth (Depth Anything v2) and segmentation (SAM 2) signals into LLaVA-1.5, then applying GRPO with a mirror-flip spatial consistency reward. Evaluated on VSR, What's Up, and CV-Bench.

Languages Python · Java · Bash · PowerShell ML/DL PyTorch · TensorFlow · Hugging Face · CNNs · Transformers (BERT, LLaMA-2) · LoRA/PEFT · FSDP · FlashAttention · Mixed precision Vision YOLOv8 (detection & segmentation) · OpenCV · ResNet · image captioning · multi-modal PET/MRI MLOps Docker · Kubernetes · Helm · ArgoCD · GitHub Actions · Prometheus · Grafana · Trivy Cloud/Compute Azure · CUDA · multi-GPU training Eval W&B · BLEU/ROUGE/BERTScore · beam search