I build reliable machine-learning systems for engineering and scientific problems, from model development and evaluation through deployment, observability, and failure analysis.

My work currently spans agentic RAG, LLM adaptation and evaluation, multimodal VLM evaluation, time-series diagnostics, recommendation systems, scientific ML, and production-oriented ML systems.

My aerospace background strongly influences how I approach AI: model capability matters, but so do provenance, evaluation integrity, safe failure behavior, reproducibility, and knowing when a system should not make a claim.

Portfolio · LinkedIn · Email

Agentic RAG · LLM Evaluation · ML Systems

Evaluation-first knowledge system over 3,233 citation-preserving NASA NTRS chunks.

• Built hybrid BM25 + dense retrieval, Reciprocal Rank Fusion, cross-encoder reranking, pgvector, evidence-sufficiency gating, and application-controlled citations.
• Adapted Qwen3-0.6B with PEFT/LoRA and developed protected evaluation for answerability, grounded refusal, citation validity, evidence support, and semantic claim coverage.
• Built a stateful LangGraph agent with schema-constrained planning, registered tool routing, bounded execution, checkpointing, human-in-the-loop review, fault injection, and safe failure handling.
• Decomposed the system into Agent, Retrieval, and Inference FastAPI services with Docker Compose, PostgreSQL/pgvector, OpenTelemetry, Prometheus, retries, health/readiness checks, and safe degradation.
• Developed TRL/GRPO grounded-agent experiment infrastructure with multi-objective rewards, anti-reward-hacking gates, leakage controls, and paired Base/LoRA/GRPO evaluation contracts.

Multimodal AI · VLM Evaluation · Reliability Engineering

Human-aligned evaluation tooling for synthetic aerospace inspection imagery.

• Designed a frozen 48-scenario benchmark with a protected 36/12 development-test split.
• Materialized 48 deterministic synthetic assets with generation metadata and SHA-256 provenance.
• Built typed evaluation, annotation, prompt and response contracts with strict protected-test and metadata validation.
• Integrated an optional local MLX-VLM development runner with exact asset/scenario/rubric binding.
• Automated validation with Ruff, strict mypy, pytest and GitHub Actions.

Recommender Systems · Production ML

Leakage-aware recommendation and serving over Microsoft MIND-small.

• Chronological train/validation evaluation with DuckDB and TF-IDF personalization.
• Training-only popularity fallback recovered all content-model abstentions.
• Final held-out results: NDCG@10 0.3664, Recall@10 0.5955, Hit Rate@10 0.6762.
• FastAPI, non-root Docker/Compose, readiness/tracing/latency telemetry, multi-platform GHCR publishing, and 400+ tests.

Scientific ML · Uncertainty · Efficient Inference

Reproducible aircraft-design benchmark linking surrogate models to uncertainty-aware engineering decisions.

• Compared Ridge, Random Forest, and HistGradientBoosting surrogates on a 6,000-sample deterministic benchmark.
• Best model reached 0.9952 mean held-out R².
• Added tree-quantile and split-conformal uncertainty plus a feasibility classifier with 99.89% ROC AUC and 2.12% false-safe rate.
• Evaluated 20,000 optimization candidates, extracted 47 Pareto designs, and validated selected designs against the physics model.
• Exported models to ONNX with equivalence tests; batch-1 surrogate latency improved from 13.795 ms to 0.305 ms.

Graduate Research Assistant · May 2025 - Aug 2026

• Built a leakage-resistant failure-detection workflow over 15,120 time-series thrust simulations, achieving 96.8% accuracy / 0.986 ROC AUC on unseen geometries.
• Contributed to the Delta Air Lines-sponsored HERO project on source-aware retrieval for analyst-led safety and risk assessment.
• Built a GREET lifecycle surrogate and integrated it with economic and transportation models for the U.S. Endowment-sponsored GREEN TEA project.

Machine Learning Engineer · Oct 2023 - Apr 2025

• Developed Python diagnostics, anomaly-detection and predictive-maintenance workflows for multivariate aircraft-engine sensor data.
• Built Azure Databricks data pipelines and engineered 1 Hz / 10 Hz time-series features.
• Worked with lifecycle engineers to validate failure modes and convert model signals into proactive maintenance insights.

Languages & data: Python, SQL/MySQL, C++, MATLAB, pandas, NumPy, SciPy, DuckDB, PostgreSQL/pgvector, Azure Blob Storage, Azure Databricks

GenAI & ML: Hugging Face Transformers, PEFT/LoRA, TRL/GRPO, LangGraph, LangChain, RAG, BM25, dense retrieval, RRF, cross-encoder reranking, Sentence Transformers, PyTorch, TensorFlow, scikit-learn, MLX-LM/MLX-VLM

ML systems: FastAPI, REST APIs, Docker/Compose, Google Cloud Run, GitHub Actions, Prometheus, OpenTelemetry, CI/CD, distributed tracing, health/readiness contracts, fault injection

Scientific ML & evaluation: surrogate modeling, uncertainty quantification, conformal prediction, constrained optimization, physics validation, held-out evaluation, bootstrap analysis, failure analysis

Reliable AI systems that combine grounding, agentic tool use, model evaluation, failure recovery, multimodal evaluation, post-training experimentation, and reproducible deployment.