I build LLM systems for people who cannot afford a wrong answer: clinicians, counselors, and field teams in low-resource settings. Most of my code falls into one of four stages, and I keep a repo for each.
| retrieve & ground | model | evaluate | decide |
|---|---|---|---|
|
medrag-toolkit Medical question answering over PubMed, OpenFDA and RxNorm. Hybrid dense + BM25 retrieval. Every sentence is checked against its source; anything unsupported is dropped, and it abstains rather than guesses. |
PlanMoE Mixture of experts over room-constraint graphs, INT8 quantized, on HF Spaces. sepsis-prediction BiLSTM that flags ICU sepsis risk 6 to 12 hours early. skiavox Chest X-ray ensemble with Grad-CAM and DICOM in. mindscope Mental health NLP benchmark, 15+ conditions, BERT and Claude baselines. |
ragcheck
Scores retrieval, faithfulness, relevance and citation coverage separately, and prints the exact sentences a RAG answer made up. Offline judge for CI, LLM judge for depth, a |
abtest-kit Power analysis, z and Welch tests, Bayesian expected loss, always-valid sequential tests, CUPED. Each method is tested against an independent reference. The peeking demo: naive t-test 32% false positives, mSPRT 1.8%. |
| when | what |
|---|---|
| Sep 2026 | Published ragcheck and abtest-kit. Both green on CI across Python 3.10 to 3.12. |
| Aug 2026 | medrag-toolkit and PlanMoE went public. Rebuilt the portfolio site on Next.js. |
| Mar 2026 | Joined the Center for Global Mental Health Equity at GWU as an AI research engineer. The agent pipeline I shipped scores counselor competency across 26K+ transcripts in 25 countries and cut per-session review from 45 minutes to 4. |
| 2025 – 2026 | ML engineer at Data Science for Sustainable Development: NLP extraction over sustainability reports and ARIMA/LSTM/XGBoost forecasting on 200K+ humanitarian records across 18 countries. |
| Jan 2025 | Started the M.S. in Data Science at George Washington University. GPA 3.75, finishing Dec 2026. |
| 2023 | Drone object detection and predictive maintenance at Asteria Aerospace, Bangalore. |
Comparing GPT-4 and Claude against human-annotated ground truth across 14 languages turned up a 23% calibration gap on low-resource languages. A RAG layer grounded in WHO EQUIP/ENACT rubrics, with reranking and hallucination flagging, brought the factual error rate down by about 40% on held-out multilingual benchmarks. Those findings feed a joint study with OpenAI on clinical LLM deployment. ragcheck is the open-source distillation of that measurement work.
Plus the parts without icons: LangChain and LangGraph, the OpenAI and Anthropic APIs, FAISS, Pinecone, Qdrant and pgvector, RAGAS, DeepEval and LangSmith, Hugging Face, spaCy, XGBoost and LightGBM, PySpark, Airflow, BigQuery, MLflow, Streamlit.
These cards are rendered by scripts/build_stats.py in this repo, refreshed every morning by a GitHub Action. No third-party widget service.
portfolio · resume · linkedin · nikhil.obuleni@gwu.edu
Open to AI engineer and data scientist roles from 2027. If you work on AI for health and want a second pair of eyes on an evaluation setup, my inbox is open.