AI Engineer
PDF ↓Ship LLM features that are grounded, measurable and hard to break.
Roles I'm targeting06
The roles below are where I want to work next. Each one links to the projects and internships that show the skills it needs.
Machine Learning Engineer
PDF ↓Train models that beat a real baseline, then run them in production.
Data Scientist
PDF ↓Design honest experiments and find out what the data can support.
Data Engineer
PDF ↓Build pipelines that ingest messy sources reliably and reproducibly.
Data Analyst
PDF ↓Turn raw logs and metrics into clear charts and decisions.
AI / ML Researcher
PDF ↓Run rigorous experiments, find the real limits, and report them honestly.
Featured projects03
Academy of
Testers
Testers
academyoftesters.comAdaptive SAT engine (BKT + IRT) and a RAG essay grader cutting AP scoring error ~23%.Open case study
Seismic
SoCal
SoCal
seismicsocal.duckdns.orgLive earthquake detection on 10 SeedLink stations: CNN→Transformer + GNN, push alerts.Open case study
BearLM
fully local RAGZero-cost cited Q&A over ~10K Berkeley CS/DS chunks. recall@1 56% → 82%.Open case study
Experience06
LBLLawrence Berkeley National LaboratoryData Science & ML Research Intern · PatchTST forecasting of XCache traffic, up to 24% lower RMSE than persistence.CBUCalifornia Baptist University · SeismicSoCalML Research Intern · diagnosed an inverse-FEA data ceiling; 0.992 ROC-AUC detection; live alert daemon.NITNumIsTokenFull-Stack Software Engineering Intern · resumable redemptions; serialization tool; load time ~2–3 s → ~300 ms.KGThe Kigumi Group · KiguLabBackend Software Engineer Intern · coverage 90% → 98%; fixed recurring AI assistant pipeline failures.BLMBearLMData & AI Engineer · local hybrid-search RAG; RAGAS faithfulness 0.55 → 0.83.AOTAcademy of TestersML & AI Software Developer · adaptive learning engine + RAG essay grading.
About
I like problems where the honest answer matters more than the impressive one.
Across forecasting, seismology, retrieval and ed-tech, my work follows the same loop: build a reproducible pipeline, pick a baseline worth beating, evaluate on data the model has never seen (chronological splits, held-out sets, ablations), and then ship the thing as a real product with a real user surface.
Sometimes the honest result is a ceiling. At CBU I showed an inverse-FEA predictor was stuck because of the data, not the model, and that redirected the team toward fixing the inputs instead of tuning models that had nothing left to find.
Toolkit
ML & deep learningPyTorch, scikit-learn, XGBoost, CNN / Transformer / GNN, PatchTST, neuralforecast
LLMs & retrievalRAG, LangChain, Ollama, Llama 3.1, embeddings, BM25 + vector hybrid search, rerankers, RAGAS
DataPandas, ETL pipelines, EDA, rolling CV, Matplotlib, Seaborn, Plotly, SQL
BackendPython, FastAPI, Java / Spring Boot, PostgreSQL, Flyway, SQLite, Chroma
FrontendReact, TypeScript, Vite, Tailwind, Capacitor (Android), hand-built SVG
OpsDocker, systemd, Caddy, Oracle Cloud, Render, Vercel, FCM push