This repository contains the portfolio site I use to explain my machine learning work through problems, decisions, experiments, and results. The site focuses on scientific machine learning, retrieval and ranking, recommendation systems, GenAI evaluation, distributed training, high-throughput LLM serving, and model deployment.
The experience section and downloadable CV include my work as a Machine Learning Engineer at Rolls-Royce from July 2023 to April 2025. I entered Georgia Tech's Aerospace Engineering PhD program, transitioned to the MS, and completed the degree in August 2026.
A project card or technology list does not explain why a model was chosen, what failed, or how a result was measured. Each case study therefore follows the same basic story:
- the problem and its constraints;
- the approach and the reason behind it;
- the evaluation or engineering evidence;
- the result, including negative findings; and
- the limits of what the project establishes.
The homepage introduces the common thread across the work. Separate pages cover AIRFAANS, AeroRAG-X, IntegrityBench, NewsLens, EdgeGenBench, AeroSynth-Eval, rocket-motor failure detection, equity backtesting, Atlanta mobility resilience, and GREEN TEA.
- AIRFAANS now reports a matched three-seed comparison of a pointwise MLP, MeshGraphNet-style GNN, and point neural operator across all 200 official AirfRANS interpolation test meshes per treatment. No architecture wins every field and force metric; Reynolds/AoA OOD and uncertainty runs remain pending.
- IntegrityBench keeps every candidate blocked. A 20,000-row mixed Civil Comments/BeaverTails experiment reached 2.29% false acceptance in-domain but 9.60% on human-reviewed ToxicChat, making the shift gap visible rather than hiding it behind the in-domain result.
- EdgeGenBench now leads with its public NASA DASHlink flight-anomaly track. The generated aircraft-design surrogate remains useful deployment evidence, but it is no longer presented as measured-aircraft model accuracy.
- AeroSynth-Eval includes a ten-seed AGDD transfer study and 1,735 GenAI-Bench preference votes. It now has a verified acquisition path for DLR's MIT-licensed 6,000+ image aircraft-dent dataset; no DLR model result is claimed until its archive and splits are audited.
- AeroRAG-X evaluates retrieval on QASPER and SciFact human annotations, audits 20,283 public TREC RAG relevance judgments and 2,840 citation-support judgments, and rejects 200/200 deliberately wrong NASA source IDs. The 50-case aerospace author audit is still pending.
- Atlanta Mobility Resilience uses 50 Census tract centroids, 101 mapped essential facilities, and MARTA's public static GTFS schedule. A fixed 20-case MARTA planner review found that the first router misses walking transfers between nearby stops. The schedule percentages are development outputs, not portfolio results, until that routing gap is fixed and retested.
- Surrogate Model Learning confirms a frozen uncertainty method on 1,030 public UCI concrete measurements. Test R² is 0.9023, and normalized 90% intervals cover 95.83% of the held-out mixture groups. A high-age tail did not trigger the distance guard more often, so the guard is not overstated.
The portfolio is a static site built with HTML and CSS. It uses no application
framework or build step, which keeps the deployed site small and makes every
page easy to inspect. Shared case-study styling lives in case-styles.css, and
project visuals and the downloadable CV live under assets/.
The update path from repository evidence to case study and resume is shown in the site architecture.
python3 -m http.server 8000Open http://localhost:8000 in a browser. Serving the files over HTTP avoids
browser restrictions that can appear when pages are opened directly from disk.
The portfolio summarizes results from the linked repositories. The repositories, frozen reports, tests, and source code remain the evidence of record. The site does not turn planned work into completed work or present synthetic fixtures as human, production, device, or safety evidence.