Curriculum Vitae

Founder and sole developer

Polars · DuckDB · PyTorch · XGBoost · ONNX Runtime · React · Tauri · FastAPI

• Shipped a local-first desktop analytics app in React and Tauri over a FastAPI backend that lets researchers ask questions of their spreadsheets in plain English, computed entirely on the machine, so data covered by IRB or FERPA never leaves it. In use by a Vanderbilt principal investigator and by a non-technical business owner.
• Routed analytical questions through exact DuckDB SQL aggregates rather than the model, and badge a result as verified only when a deterministic query produced it, which removes a whole class of numeric hallucination.
• Built an end-to-end AutoML and training pipeline: a time-budgeted XGBoost hyperparameter search and PyTorch Lightning training, with leakage detection, ONNX export, and global explanations surfaced in the interface.
• Wrote a hardware-aware inference backend that detects the GPU architecture across NVIDIA CUDA, DirectML for AMD and Intel Arc, Apple Metal, and CPU, then reconciles predictions against bound ONNX providers, so a session cannot silently fall back to CPU without saying so.
• Measured link discovery across a 2,950 file, 21.3 GB public corpus: precision 0.910 and recall 0.763 over 48 labelled groups and 86 labelled pairs.

Independent research project

Python · Evo 2 · Google Cloud · GenBank · IGV · Geneious

• Built a command-line tool and cloud-orchestration pipeline that computes per-position Shannon entropy for a DNA sequence in a single forward pass through the Evo 2 genomic language model, replacing the lab’s previous alignment-based conservation workflow.
• Automated the whole Google Cloud GPU virtual machine lifecycle, provisioning, health check and teardown, so a run costs only the minutes of GPU time it uses, and exports bedGraph, wig, GFF3 and TSV tracks that open directly in IGV, Geneious, SnapGene and Benchling.
• The GPU is rented inside the researcher’s own cloud project, so no server of mine is in the path and the bill and the sequences stay theirs. Used independently by lab researchers on their own sequences.

Software Engineer

TypeScript · Tauri · Rust · Node.js · Vitest · Electron · Flask

• Built a pixel-accurate Windows, macOS and web emulator of the AB-991EX scientific calculator: a dependency-free TypeScript core whose parser and rasteriser drive a 192 by 63 dot framebuffer across the calculator’s 12 calculation modes, on decimal.js precision matched to the hardware.
• Pinned the rendering behind 244 golden screen files, inside a suite of 3,560 tests, so any change that moves an LCD pixel has to ship the picture it now produces.
• Built the desktop telemetry backend: batched analytics, error ingest, per-IP and per-install rate limiting, and the auto-update feed.
• Automated a national math competition end to end in a Python and Electron app: Excel-driven 60-question exam generation with a synced answer key, percentile grading, and an individualised report card and certificate for every student, covered by 51 tests.

Lead Researcher

Python · scikit-learn · XGBoost · Keras · MATLAB

• Led a 30-person research team investigating engineering identity in STEM students across data collection, classification and predictive modelling, and built the internal Python tooling, data-cleaning scripts and onboarding material the team depended on.
• Rebuilt both original pipelines, a MATLAB wisdom-of-the-crowd ensemble and a Keras neural network, and scored them on the same cross-validation folds as a set of new models, so the original results could be audited rather than cited. Under repeated five-by-five cross-validation the original network reaches an R-squared of 0.281 plus or minus 0.027, against 0.417 plus or minus 0.040 for the best new model, a stacked ensemble.
• Found the study’s largest effect by removing it: seven survey items that closely proxy the target score were doing most of the work, and taking them out drops the best model from 0.417 to 0.294 plus or minus 0.030. Predicting the end-of-semester score from start-of-semester answers alone reaches only 0.115 plus or minus 0.051, which is the honest version of the task.
• Automated key data-pipeline segments in Python, including keyword and synonym detection across 27,000+ open-ended definitions and error detection over 9,000+ concept maps, saving the team 500+ hours, and restructured the team’s optimised Excel workflows.

“Using an Artificial Neural Network to Predict Engineering Identity Shifts in STEM Students.” Manuscript in preparation.

Commercial-grade syringe pump

Fusion 360 · 3D printing · embedded systems · C++ · Arduino · AccelStepper

Built a commercial-grade syringe pump from scratch using Fusion 360, 3D printing and custom hardware wiring, and wrote and deployed the embedded C++ firmware for precise flow control, an I2C LCD display and the hardware safety interlocks.