I build ML systems that prove their numbers: recommenders scored against honest baselines, fraud models thresholded in dollars, and RAG services that cite their sources or say "I don't know".
| ⏳ 4+ years AI/ML engineering |
🏢 Meta & Wells Fargo social media and banking |
📦 3 OSS repos 88 tests, all green |
💵 $3M+ fraud identified annually |
🎓 M.S. Business Analytics Anderson University, 2025 |
Not a logo wall. This is where my time has actually gone:
| Level | Stack | Signal |
|---|---|---|
 |
Python · PyTorch · scikit-learn · XGBoost · LightGBM · SQL · PySpark | Daily drivers across both roles; modeling, feature engineering, and evaluation live here |
 |
RAG · LangChain · Transformers · Hugging Face · TensorFlow · Kafka · Spark Structured Streaming · Airflow · MLflow · Docker | LLM applications, streaming inference, and the MLOps spine of my repos |
 |
Kubernetes · GitHub Actions · Jenkins · Prometheus · Grafana · Evidently AI · Power BI · Hive · AWS (EMR, EC2, S3) · Azure | Deployment, monitoring, and reporting around the models |
At Meta, building AI/ML for systems that serve billions of interactions:
- Recommendation and NLP models with PyTorch, Transformers, and Hugging Face, tuned through offline and online evaluation
- RAG and LLM applications for semantic search, summarization, and conversational AI
- Real-time streaming inference on Kafka and Spark Structured Streaming
- Model monitoring across drift, latency, and data quality with Prometheus, Grafana, and Evidently AI
Every repo below installs with pip, runs offline, and prints the numbers its README claims.
| Project | What it proves | Headline number |
|---|---|---|
| grounded-rag-service | RAG with cited answers, a faithfulness gate, and abstention on a measured curve; a chunking by retriever grid shows why retrieval metrics alone mislead | False answers: 100% ungated, 13% at the default gate, 0% strict, on 400 questions |
| fraud-detection-pipeline | Fraud scoring thresholded in dollars, not F1: PySpark feature parity, from-scratch PSI drift monitor, Airflow drift gate | Operating cost down 30.5%, false positives down 91.5% vs the default threshold |
| hybrid-movie-recommender | Hybrid recommender with a per-user confidence-weighted blend, judged against honest popularity and kNN baselines | Cold-start P@10 up 80%; 371 req/s at 20 ms p95 under load |
Wells Fargo taught me that a model's real metric is denominated in dollars: a fraud false positive is a frozen card and an angry customer, so I learned to pick thresholds by cost, not F1. Meta taught me evaluation at scale: nothing ships on vibes, everything ships on measured lift. My repos are those two lessons fused and open-sourced: the fraud pipeline prices its threshold in dollars, the recommender publishes its baselines, and the RAG service treats "I don't know" as a first-class, measured answer. In between the two jobs I finished an M.S. in Business Analytics, and took the Meta seat before the degree was even done.
Built like my services: every answer cited, every number measured.